Daily Papers

Reinforcement Learning with Verifiable Rewards (RLVR) has substantially improved Mixture-of-Experts (MoE) LLMs, but its outcome-level rewards provide only sparse supervision. On-Policy Distillation offers dense token-level feedback, yet requires a stronger compatible teacher, while on-policy self-distillation relies on hand-designed privileged information. We observe that the MoE architecture itself offers a way to resolve this dilemma, as increasing the activation budget per token exposes latent expert capacity and yields a potentially stronger teacher within the model. However, naive expansion degrades performance because newly co-activated experts are not optimized to work together. We therefore propose MoEvolve, a bidirectional distillation framework following a recovery-then-improvement strategy: we first restore the activation-expanded high-k model and further strengthen it by exploiting its additional active capacity, then reverse the direction of distillation to transfer the enhanced capabilities back into the original low-k configuration, retaining its inference-time activation budget. Across diverse MoE models and benchmarks, MoEvolve improves the initial models by up to 22.2 domain-average points and outperforms same-budget RLVR baselines by up to 6.5 points. Notably, the final low-k models often surpass their activation-expanded teachers.

Mixture-of-Experts (MoE) decouples model capacity from per-token computation, yet their scalability remains limited by the physical dimensions of depth and width. To overcome this, we propose Mixture of Universal Experts (MOUE),a MoE generalization introducing a novel scaling dimension: Virtual Width. In general, MoUE aims to reuse a universal layer-agnostic expert pool across layers, converting depth into virtual width under a fixed per-token activation budget. However, two challenges remain: a routing path explosion from recursive expert reuse, and a mismatch between the exposure induced by reuse and the conventional load-balancing objectives. We address these with three core components: a Staggered Rotational Topology for structured expert sharing, a Universal Expert Load Balance for depth-aware exposure correction, and a Universal Router with lightweight trajectory state for coherent multi-step routing. Empirically, MoUE consistently outperforms matched MoE baselines by up to 1.3% across scaling regimes, enables progressive conversion of existing MoE checkpoints with up to 4.2% gains, and reveals a new scaling dimension for MoE architectures.

Self-attention is central to Transformer performance and is often the most expensive part of the Transformer at long context lengths because its pairwise token interactions scale quadratically with sequence length. Standard dense attention also applies the same set of attention heads to every token regardless of token difficulty or information content. This uniform activation can waste compute, especially as sequences grow longer and attention cost increases rapidly. We propose Grouped Query Experts (GQE), a mixture-of-experts layer on top of grouped-query attention (GQA). Within each GQA group, a router selects k query-head experts per token while all key-value (KV) heads remain dense and unchanged. Thus, GQE keeps the KV cache benefits of GQA and reduces only the active query-head computation. On a fixed 30B token budget at the 250M parameter scale, GQE matches the all-active GQA baseline in downstream accuracy while activating half the query heads per token.

When a large language model (LLM) agent executes the same task, token consumption can vary by over an order of magnitude across runs. The agent chooses its next steps based on tool feedback and intermediate results, while the growing context steadily inflates the input size of every subsequent call. The total consumption of a task is therefore hard to predict before execution and the prediction must be revised as the run unfolds. In this paper, we propose TokenCast, which learns a composable cost representation for each execution segment, recording its own consumption and the context growth it introduces. Composing adjacent segments yields a cumulative estimate that captures the extra input cost incurred when context from earlier segments is re-read by every later call. As execution unfolds, newly observed evidence refreshes the forecast, requiring no additional LLM calls and incurring a mean cumulative prediction time of 32.8 ms per run on SWE-bench Verified. Across 4 task suites and 6 agent models, TokenCast's mean absolute error reduction against the strongest comparator averages 14.5% over 96 evaluated combinations. In offline budget-control replay, TokenCast uses 21.3% fewer tokens on average than a fixed-budget policy at matched trace completion. The code is available at https://github.com/DEFENSE-SEU/TokenCast.

Reinforcement learning with verifiable rewards (RLVR) generates hundreds of thousands of tokens per training step, with rollout generation dominating the computational cost. The overall token budget can be controlled along two main dimensions: (i) deciding which prompts to allocate rollouts to, and (ii) deciding how long each rollout should be. Prior work has generally controlled only one of these dimensions at a time. We show that jointly tuning both decisions under a shared compute budget improves both reasoning quality and wall-clock training time. We instantiate this view as DUal-controlled tokEn allocaTion (DUET), a computationally efficient layer over GRPO that uses a lightweight pre-rollout surrogate of prompt informativeness to set how many rollouts each prompt receives, and a marker-gated abort rule with importance reweighting to set when to stop them. On Qwen3-1.7B trained on MATH, DUET outperforms full-budget GRPO and the other three budget-aware baseline methods. DUET's advantage further generalizes to other benchmarks across math and coding, and is on par with the best baseline on the scientific Q\&A domain, while also achieving a 1.62times wall-clock speedup. More notably, using only 50\% of the token budget, DUET still outperforms all baseline methods at their full budget, achieving an even higher 2.51times speedup over full-budget GRPO. We verify the high performance of DUET on other backbone LLMs, including Qwen3-4B and Llama-3.2-3B-Instruct. Notably, the gap between DUET and the strongest baseline widens as the budget tightens, contrary to the usual pattern in which efficient methods trade off quality as compute decreases. More broadly, these results suggest that DUET budget-aware control strategies are valuable not only for accelerating training, but also for improving the quality of the learning signal.

Serving large language models at long context is bottlenecked by the key-value (KV) cache, which is read in full at every decode step. Attention keys are locally low-rank though globally high-rank: a fixed low-rank sketch shared across pages is provably blind to page-specific directions, while at the same summary size a page's own basis ranks pages and keeps carriers far better. LOCKS gives every page its own rank-r spectral summary (resident, a tenth of the cache at r{=}8 and a twenty-fifth at r{=}2), reconstructs within-page logits, estimates each page's attention mass by log-sum-exp, and attends only the top pages; selection itself reads no candidate keys or values. Selecting on this summary alone stays within about a point of the full cache on long-document QA (LongBench-v1; Llama-3.1-8B), tracks the read-every-key exact-LSE oracle on retrieval-dense RULER down to the smallest budgets, and holds quality furthest into the small-budget regime on long-form reasoning (AIME26, MATH-500; Qwen3-4B), where selectors and eviction-based reasoning compressors both fall away. At a 2048-token budget LOCKS matches FullKV aggregate quality at 100K+ context (GLM-4-9B-Chat-1M) while attending about 2% of the tokens; since the summary is scanned in full each step, the per-step KV read falls by about 10--25times across that rank range, and this halves per-token decode latency (2.0times at 1M tokens on one H200 NVL, measured at r{=}8) against dense attention. LOCKS ships as a drop-in plugin for unmodified vLLM, with batched decode running in full CUDA graphs.

Replacing one LLM agent with a collaborating team can raise accuracy, but whether scaling the team helps, and which architecture to scale, is unclear. Sweeping eight agent orchestration architectures across five instruction-tuned 7-9B models, five short-answer benchmarks, and an executable-code benchmark up to 30 calls, we find that the returns to team scaling are sharply task-dependent: from three to thirty calls accuracy rises by up to 17 points on the two arithmetic word-problem benchmarks (GSM8K, GSMHard) but by at most four on ARC, GPQA, and MMLU, for every architecture, a split the usual task-averaged number conceals. Proposer-Critic captures the arithmetic gains, scaling steepest and, in aggregate, surpassing every other architecture at the largest budget (item-clustered intervals exclude zero), though it ranks among the weakest elsewhere, and no architecture wins across tasks. We explain these trajectories with an exact generate-transform decomposition. Partitioning any workflow into proposal coverage and a downstream transform, any accuracy change splits exactly into an extensive coverage dividend and an intensive transformation change. The decomposition diagnoses each task: arithmetic offers coverage headroom that a critic-guided transform converts, whereas the multiple-choice benchmarks either saturate in coverage or fail to convert it, and on open-ended code generative recovery nearly vanishes so accuracy tracks coverage. At equal call budgets token cost still varies 2.1x. Extra calls therefore create candidate opportunity that only some architectures, on some tasks, convert. Team scaling is a task- and architecture-specific bet, not a uniform lever.

Agentic AI development today runs on token maxing: buying capability with tokens -- longer reasoning traces, more turns, wider tool payloads, bigger replayed contexts -- so tokens per task grow faster than task value. Falling per-token prices mask the pattern; total spend rises anyway. We argue the decisive lever against token maxing is the harness: the orchestration layer that assembles context, exposes tools, sequences turns, delegates work, and carries enterprise observability and governance. We isolate it with a controlled swap: 22 locked evaluation tasks, six foundation models (Claude Sonnet 4.6, Gemini 3.1, Gemini Flash 3.5, Qwen 3.6, GLM 5.1, Palmyra X6), changing only the orchestration layer -- a frozen conventional production loop versus the Writer Agent Harness. Holding models constant, the harness cuts blended cost per task 41% (0.21->0.12), median wall-clock 44% (48s->27s), and tokens per task 38% (14.2k->8.8k), with task-completion quality at parity (0.78->0.81, directional at this sample size). Efficiency is model-invariant -- every model gets cheaper (33-61%) -- while quality gains are capability-dependent: a model's gain correlates almost perfectly with its baseline strength (r=0.99, n=6), a phenomenon we term harness leverage. Quality per dollar rises 82%; task-completions per million tokens rise from 54.9 to 92.0. On this workload the orchestration layer moved cost per task more than the full spread of the model menu did. We formalize token economics at the orchestration layer (including effective input price under prompt caching), detail the six mechanism families behind the effect -- cache-shape discipline to failure-spend governance -- compare six widely used agent systems on the same axes, and argue the harness is the one component whose efficiency multiplies across every model an organization runs -- present and future.

Mixture-of-Experts (MoE) has become a dominant architecture for scaling large language models due to their sparse activation mechanism. However, the substantial number of expert activations creates a critical latency bottleneck during inference, especially in resource-constrained deployment scenarios. Existing approaches that reduce expert activations potentially lead to severe model performance degradation. In this work, we introduce the concept of activation budget as a constraint on the number of expert activations and propose Alloc-MoE, a unified framework that optimizes budget allocation coordinately at both the layer and token levels to minimize performance degradation. At the layer level, we introduce Alloc-L, which leverages sensitivity profiling and dynamic programming to determine the optimal allocation of expert activations across layers. At the token level, we propose Alloc-T, which dynamically redistributes activations based on routing scores, optimizing budget allocation without increasing latency. Extensive experiments across multiple MoE models demonstrate that Alloc-MoE maintains model performance under a constrained activation budget. Especially, Alloc-MoE achieves 1.15times prefill and 1.34times decode speedups on DeepSeek-V2-Lite at half of the original budget.

The wide adoption of AI agents in complex human workflows is driving rapid growth in LLM token consumption. When agents are deployed on tasks that require a significant amount of tokens, three questions naturally arise: (1) Where do AI agents spend the tokens? (2) Which models are more token-efficient? and (3) Can agents predict their token usage before task execution? In this paper, we present the first systematic study of token consumption patterns in agentic coding tasks. We analyze trajectories from eight frontier LLMs on SWE-bench Verified and evaluate models' ability to predict their own token costs before task execution. We find that: (1) agentic tasks are uniquely expensive, consuming 1000x more tokens than code reasoning and code chat, with input tokens rather than output tokens driving the overall cost; (2) token usage is highly variable and inherently stochastic: runs on the same task can differ by up to 30x in total tokens, and higher token usage does not translate into higher accuracy; instead, accuracy often peaks at intermediate cost and saturates at higher costs; (3) models vary substantially in token efficiency: on the same tasks, Kimi-K2 and Claude-Sonnet-4.5, on average, consume over 1.5 million more tokens than GPT-5; (4) task difficulty rated by human experts only weakly aligns with actual token costs, revealing a fundamental gap between human-perceived complexity and the computational effort agents actually expend; and (5) frontier models fail to accurately predict their own token usage (with weak-to-moderate correlations, up to 0.39) and systematically underestimate real token costs. Our study offers new insights into the economics of AI agents and can inspire future research in this direction.

Recent deep-thinking large language models often reason extensively to improve performance, but such lengthy reasoning is not always desirable, as it incurs excessive inference costs with disproportionate performance gains. Controlling reasoning length without sacrificing performance is therefore important, but remains challenging, especially under tight thinking budgets. We propose budget guidance, a simple yet effective method for steering the reasoning process of LLMs toward a target budget without requiring any LLM fine-tuning. Our approach introduces a lightweight predictor that models a Gamma distribution over the remaining thinking length during next-token generation. This signal is then used to guide generation in a soft, token-level manner, ensuring that the overall reasoning trace adheres to the specified thinking budget. Budget guidance enables natural control of the thinking length, along with significant token efficiency improvements over baseline methods on challenging math benchmarks. For instance, it achieves up to a 26% accuracy gain on the MATH-500 benchmark under tight budgets compared to baseline methods, while maintaining competitive accuracy with only 63% of the thinking tokens used by the full-thinking model. Budget guidance also generalizes to broader task domains and exhibits emergent capabilities, such as estimating question difficulty. The source code is available at: https://github.com/UMass-Embodied-AGI/BudgetGuidance.

While agents are increasingly spending more resources, today agent cost is mostly measured only after execution. A Budget-Aware Agent (BAGEN) should treat budget as an active control signal, rather than a passive cost metric. We first systematically define budget estimation as internal budgets (from agent computation) and external budgets (from agent actions). We then formalize budget-awareness as progressive interval estimation: at each step of a plan, an agent should predict an upper and lower bound on remaining budget, and alert when completion is unlikely. Scoring with a rollout-replay protocol, we find consistent failure patterns on four environments and five frontier agents: (1) strong agents do not necessarily have strong budget-awareness, with correlation r=0.35. (2) frontier models are consistently over-optimistic, continue spending on tasks that are unlikely to succeed, instead of alerting the user early. (3) budget-aware signal is actionable and trainable. Early stop saves 28-64% tokens on failed trajectories, and SFT+RL strengthens early stop and alert behavior. (4) precise interval calibration remains challenging, with interval coverage capping at 47% after SFT+RL. Project page: https://ragen-ai.github.io/bagen/

Long-horizon LLM agents accumulate sessions of up to millions of tokens, far more key-value (KV) cache than one GPU holds, so serving systems keep a small budget of tokens and evict the rest. Eviction methods differ in which tokens they keep but share one goal: that the tokens a later step needs are still cached. We show this is not enough. In 100K-token coding sessions on Qwen3-8B with a 2,048-token budget, a planner-guided policy that keeps over 99% of the needed tokens resident still fails late steps increasingly often: its accuracy falls from 93% to 64% as the gap between a definition and its use grows from 2K to 90K tokens, at residency 0.996. The needed tokens are resident but unretrievable: what grows is their rotary-position distance to the query. We present PACE (Planner-Aware Context Engine), a KV-cache engine that couples plan-aware retention, in which the agent's planner tells the cache which context future steps will read, with bounded positions, which renumber kept tokens so their rotary distance to the query stays below 8K at any session length. In a pre-registered evaluation on fresh tasks, PACE passes 97% of hidden unit tests at 100K tokens, against 42% for H2O and 3% for SnapKV under the same budget, and needs both components: either alone reaches 63% (the combination beats each, p < 10^-4). The result holds with length: in 1M- and 2M-token sessions on one 24 GB GPU, with definitions written 0.9M and 1.8M tokens before use, PACE passes 99% and 100% of checks and recalls every session fact, while bounded positions without foresight falls from 33% to 13%; PACE holds under 0.7 GB of KV state where the full cache would need 147-295 GB. Renumbering is exact, costs no measurable perplexity, and adds 9% wall time. On Mistral-7B, whose coding ability caps it at 40% even with nothing evicted, PACE stays at that ceiling while retention alone falls below it.

Reasoning is critical for large language models (LLMs) to excel in a wide range of tasks. While methods like Chain-of-Thought (CoT) reasoning enhance LLM performance by decomposing problems into intermediate steps, they also incur significant overhead in token usage, leading to increased costs. We find that the reasoning process of current LLMs is unnecessarily lengthy and it can be compressed by including a reasonable token budget in the prompt, but the choice of token budget plays a crucial role in the actual compression effectiveness. We then propose a token-budget-aware LLM reasoning framework, which dynamically estimates token budgets for different problems based on reasoning complexity and uses the estimated token budgets to guide the reasoning process. Experiments show that our method effectively reduces token costs in CoT reasoning with only a slight performance reduction, offering a practical solution to balance efficiency and accuracy in LLM reasoning. Code: https://github.com/GeniusHTX/TALE.

LLM-based Multi-Agent (LLM-MA) systems are increasingly applied to automate complex software engineering tasks such as requirements engineering, code generation, and testing. However, their operational efficiency and resource consumption remain poorly understood, hindering practical adoption due to unpredictable costs and environmental impact. To address this, we conduct an analysis of token consumption patterns in an LLM-MA system within the Software Development Life Cycle (SDLC), aiming to understand where tokens are consumed across distinct software engineering activities. We analyze execution traces from 30 software development tasks performed by the ChatDev framework using a GPT-5 reasoning model, mapping its internal phases to distinct development stages (Design, Coding, Code Completion, Code Review, Testing, and Documentation) to create a standardized evaluation framework. We then quantify and compare token distribution (input, output, reasoning) across these stages. Our preliminary findings show that the iterative Code Review stage accounts for the majority of token consumption for an average of 59.4% of tokens. Furthermore, we observe that input tokens consistently constitute the largest share of consumption for an average of 53.9%, providing empirical evidence for potentially significant inefficiencies in agentic collaboration. Our results suggest that the primary cost of agentic software engineering lies not in initial code generation but in automated refinement and verification. Our novel methodology can help practitioners predict expenses and optimize workflows, and it directs future research toward developing more token-efficient agent collaboration protocols.

Recent advancements in Large Language Models (LLMs) have leveraged increased test-time computation to enhance reasoning capabilities, a strategy that, while effective, incurs significant latency and resource costs, limiting their applicability in real-world time-constrained or cost-sensitive scenarios. This paper introduces BudgetThinker, a novel framework designed to empower LLMs with budget-aware reasoning, enabling precise control over the length of their thought processes. We propose a methodology that periodically inserts special control tokens during inference to continuously inform the model of its remaining token budget. This approach is coupled with a comprehensive two-stage training pipeline, beginning with Supervised Fine-Tuning (SFT) to familiarize the model with budget constraints, followed by a curriculum-based Reinforcement Learning (RL) phase that utilizes a length-aware reward function to optimize for both accuracy and budget adherence. We demonstrate that BudgetThinker significantly surpasses strong baselines in maintaining performance across a variety of reasoning budgets on challenging mathematical benchmarks. Our method provides a scalable and effective solution for developing efficient and controllable LLM reasoning, making advanced models more practical for deployment in resource-constrained and real-time environments.

Scaling test-time computation improves performance across different tasks on large language models (LLMs), which has also been extended to tool-augmented agents. For these agents, scaling involves not only "thinking" in tokens but also "acting" via tool calls. The number of tool calls directly bounds the agent's interaction with the external environment. However, we find that simply granting agents a larger tool-call budget fails to improve performance, as they lack "budget awareness" and quickly hit a performance ceiling. To address this, we study how to scale such agents effectively under explicit tool-call budgets, focusing on web search agents. We first introduce the Budget Tracker, a lightweight plug-in that provides the agent with continuous budget awareness, enabling simple yet effective scaling. We further develop BATS (Budget Aware Test-time Scaling), an advanced framework that leverages this awareness to dynamically adapt its planning and verification strategy, deciding whether to "dig deeper" on a promising lead or "pivot" to new paths based on remaining resources. To analyze cost-performance scaling in a controlled manner, we formalize a unified cost metric that jointly accounts for token and tool consumption. We provide the first systematic study on budget-constrained agents, showing that budget-aware methods produce more favorable scaling curves and push the cost-performance Pareto frontier. Our work offers empirical insights toward a more transparent and principled understanding of scaling in tool-augmented agents.

We introduce APEX-Accounting, a benchmark built by Mercor in partnership with Ramp, to assess whether frontier models can do the real work of accountants. Tasks include reconciling accounts, accruing expenses, posting transactions, and producing reports. The private eval set comprises 160 tasks, split across 10 worlds. Each world contains an accounting system, as well as spreadsheets, PDFs, and other files. Every task was authored and solved by experts in accounting and bookkeeping, who also wrote grading rubrics. Across nine frontier models, Claude-Fable-5 (Max) leads with 56.4% Mean Criteria@3, ahead of Muse-Spark-1.1 (xHigh) at 52.6%. No model scores more than 2.6% Pass^8 (GPT-5.6-Sol (Max+Pro)) and the highest Pass@8 is 21.5% (Muse-Spark-1.1 (xHigh)). We experiment with increasing the token budget from 1 to 50 and observe an instance of Simpson's paradox: scores increase as the token budget increases but within a given budget-constrained harness, scores are lower on tasks where the model spends more tokens. As APEX-Accounting is a closed benchmark, leaderboard evals can be run for any frontier model on request.

As context lengths scale, attention increasingly becomes a primary computational bottleneck in large language models. Standard Transformers remain powerful but computationally inefficient, as they allocate the same attention budget to every token regardless of its contextual demand. Existing local-global hybrids provide a more efficient alternative by mixing restricted- and full-context attention, but they typically allocate span statically across layers or heads. To address these limitations, we propose LoGo, a token-level dynamic local-global attention mechanism that uses attention span as a direct proxy for attention budget allocation. Each LoGo layer contains coupled local and global branches: all tokens receive efficient local attention over a restricted context window, while a learned gate activates global attention with full-context access only for tokens requiring long-range information. A threshold-based budget controller maintains a target global ratio without auxiliary losses, and a progressive masking schedule stabilizes training before sparse routing takes effect. We further implement query-sparse Triton kernels that convert reduced global-attention computation into practical speedups. Extensive experiments validate LoGo's effectiveness, showing that it preserves the scaling behavior of full-attention Transformers across model sizes. In controlled comparisons, LoGo improves over the full-attention Transformer and matched-budget static local-global hybrids, with clear gains on long-range retrieval. Analysis further shows that LoGo learns interpretable span allocation patterns. These results suggest that learned token-level span allocation is an effective and scalable way to improve the long-context performance-compute trade-off.

Most image generation models rely on uniform tokenization, allocating the exact same computational budget to equally-sized image patches. This static paradigm cannot adapt to different resource constraints at inference time, and yields suboptimal quality-cost tradeoff by devoting the same effort to both plain backgrounds and intricate details. We propose BudgetPix, an adaptive tokenization framework that dynamically allocates compute based on visual complexity and spatial layout, enabling flexible computational budgeting at inference time. BudgetPix comprises three key components: (1) an adaptive encoder that maps a fixed-size image to a variable-length token sequence using an entropy-guided quadtree alongside a multi-scale patch embedder; (2) a scale-aware decoder reconstructs fixed-resolution images from multi-scale token sets; and (3) a flexible training and sampling schedule that enables pixel-space denoisers to operate across variable token counts. BudgetPix seamlessly integrates with existing pixel-space diffusion architectures, enabling a single checkpoint to be operated at a wide range of compute budgets. Evaluated on text-to-image generation, BudgetPix matches the fidelity of MiniT2I-L at 512^2 and PixelDiT at 1024^2 using just 25% of the original compute budget. In class-conditional generation using a MeanFlow backbone, BudgetPix requires merely 60% of the full compute budget to produce images with near-zero quality degradation, observing a marginal 0.8-point increase in FID. Comprehensive assessments by human and VLM judges confirm that BudgetPix establishes a significantly improved quality-efficiency tradeoff over prior budget-adaptive baselines. More details are available at our project page: https://karaozgur.com/BudgetPix

Large language model (LLM) agents allocate test-time compute adaptively as they revise solutions, use tools, explore alternatives, and decide when to stop. This test-time strategy makes it difficult to measure how agent performance scales. We study open-ended tasks that provide continuous scores for intermediate submissions, making progress observable throughout long trajectories. We propose Elo-per-token analysis, which tracks the best solution found at each token budget and uses a Bradley-Terry model to aggregate within-task orderings into Elo ratings across tasks with different score scales. We apply it to four general-purpose agents on four open-ended benchmarks, with sessions of up to 100M tokens, and to three feedback-driven LLM optimization harnesses in controlled single-task interventions. Independent sampling provides a theoretically characterized reference, for which Elo grows linearly with log compute. Against this reference, agents can initially convert tokens into Elo faster than independent sampling, but their marginal gains diminish and eventually fall below the reference. In contrast, the strongest historical human contestants improve superlinearly over contest time on shared AtCoder Heuristic Contest tasks, providing evidence of continual learning and substantial headroom after agents slow down. We define the scaling inflection point as the per-session budget where marginal Elo gains match the independent-sampling reference. Using this point as the per-session budget, we split 100M tokens across parallel sessions on FrontierCS Polyomino Packing, gaining +264 Elo over one long session and +355 over ten short sessions.

Modern language models are increasingly asked to handle several tasks within a single prompt. Doing so requires solving each task correctly while integrating the resulting answers into a coherent response. In this paper, we identify a fundamental limitation in this process, where models make substantially more errors when tasks are composed, even when they can solve every constituent task reliably in isolation. This degradation cannot be fully explained by the accumulation of independent errors. Across both open-weight models and the frontier model Claude Opus 5, our findings instead point to an implicit thinking budget underlying this effect. In particular, under composition, models use substantially fewer tokens per task than in isolation, even when the full sequence remains well within the context window. Tasks that require longer solution traces in isolation are consequently more fragile under composition, i.e., models make more errors on prompts containing these tasks. In isolation, non-fragile tasks receive more reasoning than they need and waste compute. However, by grouping tasks according to their fragility, the implicit budgeting can be manipulated to reduce token usage by 43\% across a range of math and linguistic tasks while also preserving accuracy.

Coreset selection picks a representative subset of the labeled training set to make training cheaper. However, it is usually evaluated by downstream accuracy at a fixed subset size, ignoring both the time spent selecting the subset and the training recipe behind each reported number. We introduce an end-to-end benchmark that standardizes downstream training and charges selection and training to the same auditable wall-clock budget, spanning 4 datasets from CIFAR-10 to ImageNet-1K, 11 selectors, 5 fractions, and 3 seeds, with over 1,500 released runs. Repeated-sampling work has shown that budget-aware evaluation already favors random strategies. Our two budget studies test whether that verdict survives when every selector is granted its most favorable operating point. Across eight wall-clock budget anchors on each of CIFAR-10 and Tiny ImageNet, no anchor is won by a sophisticated selector: every winner is class-balanced random sampling, repeated random sampling, or full-data training. In fixed-budget duels on ImageNet-1K, training on all data for fewer epochs beats every selection strategy we probe while also costing the least. A per-dataset cost audit shows that selection cost is dominated at every scale by a fixed full-dataset scan, so it cannot be amortized away by selecting a smaller fraction, and its absolute size does not extrapolate from one dataset to another. We further quantify when selection does pay back through subset reuse, and document 9 correctness fixes to a widely used codebase, one of which shifts a standard Herding baseline by nearly 6 points. Selection time is not free preprocessing, and an evaluation that ignores it measures the wrong quantity.

AI agents may consume massive tokens across a wide range of tasks, but how many of those tokens are really necessary? Although prior work has shown that agents can incur unexpectedly high token costs, we still lack a systematic way to identify and measure token waste. In this work, we introduce a unified framework for measuring token waste across heterogeneous agent systems. Our framework records and normalizes complete agent trajectories, identifies removable spans using an LLM-as-a-Judge (LLMaJ), and maps the identified text spans back to the original trajectory for token-level measurement. We validate the identified waste through direct interventions by removing marked spans, resuming agent execution, and measuring changes in task outcomes and subsequent actions. In online pruning experiments, we remove 41-46% of the model's reasoning while largely preserving task performance across two distinct agent benchmarks. Using this framework, we analyze agents across multiple settings. Our analysis shows that token waste has a system-level structure that aggregate token usage alone cannot reveal. First, waste varies more across tasks and agent settings than across models. Second, reasoning and visible execution can exhibit different waste patterns. Third, different harnesses can introduce distinct sources of inefficiency even under the same model and task. Overall, our framework transforms token waste into an actionable optimization signal, enabling researchers and practitioners to train more efficient models, design better reasoning strategies, and optimize agent harnesses toward reducing unnecessary computation rather than merely reducing token usage.

Mixture-of-experts models make nearly trillion-parameter capacity accessible with sparse per-token computation, provided that the serving system can distribute the weights and coordinate their execution. We present Cascadia's resident execution of Inkling, a 975B-total/41B-active-parameter model, on eleven Intel Core Ultra X7 358H AI PCs, each with 64 GB of memory, Arc B390 integrated graphics and gigabit Ethernet. We contribute a custom resident MoE engine that preserves Inkling's routing rules, constructs compressed graphs for OpenVINO's fused iGPU primitives, and coordinates FP16 expert computation with FP32 output restoration. The engine fits six consecutive decoder layers per machine and represents dense feed-forward blocks as all-active expert slices, reducing measured dense-layer call time from approximately 8.1 to 4.5 ms. A streaming pipeline coordinates concurrent generation, while captured-state draft evaluation measures agreement with the deployed numerical path. Paired measurements at fifteen concurrency levels from 1 to 176 streams reach 60.29 aggregate decode tokens/s at 88 streams, with 46.87 tokens/s over the complete serving phases. At fifteen streams, median first-token latency is 6.05 s. Raising the context budget from the 1,024-position default, real prompts of 1k to 64k tokens recover the embedded code in all 19 measured answers, with first-token time growing as aN+bN^2 and decode latency growing approximately linearly, both bounded by a single-threaded CPU attention loop rather than by memory, which holds 512k positions per stream. Evaluation on captured fleet states separates the effects of vocabulary selection and weight quantization on draft agreement. Together, these contributions establish an execution and evaluation approach for large sparse models on distributed client systems with shared CPU-GPU memory.

As coding agents move from supervised code completion to unattended, around-the-clock exploration, their work expands from isolated predictions into long trajectories of reasoning, tool use, and feedback. Token efficiency therefore becomes important for scaling recursive self-improvement. We take an RSI-inspired approach at the harness layer, scaling auto-research loops across increasingly numerous and diverse environments for harness rollouts. At this scale, the process yields reusable improvements that transfer beyond their development setting, moving automated harness discovery toward production-level outcomes. Four mechanisms survive selection and form SoL-Pi, spanning action execution, context compaction, observation handling, and delegated reading. On the 51-task EdgeBench evaluation, SoL-Pi achieves performance comparable to Pi across GPT-5.6 Sol and Opus 5 while reducing recorded token traffic by 44.7-49.0% and API cost by about one third. In other words, estimated hourly savings are \8.75-13.50 relative to native Codex and Claude Code harnesses, and \4.36-5.71 relative to Pi.

The Model Context Protocol (MCP) has become a common interface for connecting large language model (LLM) agents to external tools, but its reliance on stateless, eager schema injection imposes a hidden per-turn overhead the MCP Tax or Tools Tax that practitioner reports place between roughly 10k and 60k tokens in typical multi-server deployments. This payload inflates the key-value cache, is associated with reasoning degradation as context utilization approaches published fracture points around 70%, and turns token budgets into a recurring operational cost. We introduce Tool Attention, a middleware-layer mechanism that generalizes the "Attention Is All You Need" paradigm from self-attention over tokens to gated attention over tools. Tool Attention combines (i) an Intent Schema Overlap (ISO) score from sentence embeddings, (ii) a state-aware gating function enforcing preconditions and access scopes, and (iii) a two-phase lazy schema loader that keeps a compact summary pool in context and promotes full JSON schemas only for top-k gated tools. We evaluate on a simulated 120-tool, six-server benchmark whose per-server token counts are calibrated to public audits of real MCP deployments. In this simulation, Tool Attention directly reduces measured per-turn tool tokens by 95.0% (47.3k -> 2.4k) and raises effective context utilization (a token-ratio quantity) from 24% to 91%. End-to-end figures for task success, latency, cost, and reasoning quality are reported as projections derived from the measured token counts combined with published deployment telemetry; they are not measured on live LLM agents, and we mark projected values explicitly throughout. Taken together, the results support a simple thesis: protocol-level efficiency, not raw context length, is a binding constraint on scalable gentic systems. The code for this work is accessible at https://github.com/asadani/tool-attention

Mixture-of-Experts (MoE) language models scale model ability with sparsely activated experts, making this architecture a standard recipe for modern large models. However, sparse activation does not remove the deployment burden of storing and serving all experts, and the available deployment budget can vary substantially across devices, users, and workloads. Existing MoE compression methods are still largely fixed-budget, typically optimizing one compressed endpoint at each chosen target budget. We study a different setting: converting a large pretrained MoE LLM into a nested family of deployable subnetworks across budgets. Our method first ranks expert FFN channels by their importance, then lets each expert learn a discrete action to prune its channels. By gradually increasing cost pressure, a single action-training run exports a series of action masks from high to low budgets, each of which identifies a reliable smaller subnetwork nested in the ranked base model. Moreover, we use a single recovery fine-tune at a mid pruning budget (40%) to recover degraded model quality and transfer the recovered model to other unseen budgets. Overall, our framework surpasses recent MoE compression baselines. Specifically, on Qwen2-57B-A14B, our method retains ~99.8% of base performance while pruning 50% of routed expert parameters even without fine-tuning. For deployment, our pruned subnetworks deliver real memory reduction and throughput gains, and further support realtime online budget switching with kernel-level co-design.

Generative AI is billed per token, yet tokens are not equally expensive. Each token enters the KV-cache and raises the cost of those that follow, so energy grows non-linearly along a sequence whose tokens look identical. No analytical model states what a token costs, and the most accurate SOTA model is a fit, not a law: it absorbs architecture into fitted coefficients, requiring per-LLM fitting. We derive closed-form expressions of the computational and memory-access complexity of LLM inference in terms of architectural hyperparameters, covering MHA and GQA, and compare a family of analytical models of increasing granularity. From these, we obtain the SIMPLE Inference Model to Predict LLMs Energy (SIMPLE), a four-parameter law that separates prefill from decode, and decode into KV-cache cost and next-token computation. Over traces of 13 dense LLMs from six families on NVIDIA H100 GPUs, it attains R2 = 0.969 and 11.97% MAPE, 3.1-3.5x below the SOTA, without per-LLM calibration. The SIMPLE law explains as well as predicts: prefill energy grows quadratically with the hidden dimension, so larger LLMs suffer most from long inputs; and more aggressive GQA KV grouping reduces the per-token energy by up to 3x relative to standard MHA. Increasing the reasoning token budget pushes output length beyond the analytically derived energy minimum, driving super-linear energy growth; although accuracy grows with the budget, energy grows faster, so Intelligence per Joule peaks at an intermediate budget and falls by up to 3.14x across a 4.0x reasoning-budget range. Agentic serving raises energy per request by 13.0x relative to non-agentic ones, driven by the growth of both the prefill and the KV-cache terms with input length; while prefix caching removes most of the prefill cost (-93%), the KV-cache term is unaffected, so even a cached agentic request costs 4.8x more.

Multi-head attention enables Transformers to capture diverse representations, but all attention heads are typically activated for every input, regardless of task complexity. For coarse-grained tasks such as text classification, where relevant information is often global, this fixed allocation can introduce unnecessary computation. We propose BudgetFormer, a Transformer architecture that dynamically allocates attention heads on a per-input basis. The model learns both a head budget and a relevance distribution to select the most informative heads. To support effective head selection, we introduce a training strategy that balances exploration and exploitation. Experiments on text classification tasks show that BudgetFormer reduces FLOPs and memory usage while matching or surpassing the performance of standard multi-head attention. These results highlight adaptive head allocation as an effective approach to improving Transformer efficiency and performance.

Thousands of diverse benchmarks have been developed to measure the quality of large language models (LLMs). Yet prior work has demonstrated that LLM performance is often sufficiently explained by a small set of latent factors, or abilities. This suggests the potential for more efficient and principled benchmarking, but it remains difficult to compare the quality of different methods. Motivated by predictive validity, we argue that the quality of a benchmarking framework should be grounded in how efficiently it enables the prediction of model performance on unseen tasks. To analyze this objective, we collect the "Wide-scale Item Level Dataset" (WILD), a dataset of item-model response pairs, comprising evaluations of 65 models on 109,564 unique items spanning 163 tasks drawn from 27 datasets. This dataset enables the first analysis of how different techniques can predict a model's performance on a large, diverse collection of unseen tasks under different budget constraints. We demonstrate that combining a modified multidimensional item response theory (IRT) model with adaptive item selection driven by optimal experimental design can predict performance on 112 held-out benchmark tasks with a mean absolute error (MAE) of less than 7%, and can do so after observing only 16 items. We further demonstrate that incorporating cost-aware discount factors into our selection criteria can reduce the total tokens needed to reach 7% MAE from 141,000 tokens to only 22,000, an 85% reduction in evaluation cost.

Reinforcement learning (RL) has become a key driver of progress in large language models, but scaling RL to long chain-of-thought (CoT) trajectories is increasingly constrained by backpropagation over every generated token. Even with optimized rollout engines, full-token updates can consume a large fraction of total training cost, turning token length into a hidden tax on RL. We introduce Not All Tokens Are Needed (NAT), a unified framework that makes the token budget a first-class optimization primitive. NAT updates the policy using only a selected subset of generated tokens while preserving the learning signal of full-sequence RL. The core idea is an unbiased partial-token policy-gradient estimator via Horvitz-Thompson reweighting, which ensures statistically correct gradients despite subsampling. We instantiate NAT with two simple, plug-and-play token selection schemes: Uniform Random Sampling (URS) and Random Prefix Cutting (RPC), both of which reduce forward and backward compute and memory without modifying the reward computation or rollout pipeline. Across mathematical reasoning benchmarks, NAT matches full-token GRPO performance while using as few as 50% of tokens, providing an efficient and orthogonal pathway to scaling RL beyond the limits imposed by long trajectories. In our experiments, RPC saves 18% peak GPU memory and 29% forward and backward RL training time for Qwen3-8B.

Context-reduction layers for API-based coding agents, including command-output compressors, retrieval rankers, and API-boundary proxies, are commonly evaluated by how much context or tool output they remove. We ask a different question: which interventions actually reduce end-to-end billed cost while preserving task success? Our primary evidence is a pre-specified, hash-frozen, paired campaign of 2,908 provider-billed Claude Code runs, of which 2,848 were analyzed, covering 103 tasks, seven repositories, and three models. The campaign compared a baseline with two generations of hook-based compression and an API-boundary proxy within a broader measured program of roughly 5,500 billed executions. Three findings emerge. First, prompt-cache traffic dominated cost composition, accounting for about 87% of reconstructed four-component cost (about 80% of the actual bill), with an 8.7% dollar-weighted residual not attributable from retained telemetry. Second, local payload reduction was not a reliable predictor of end-to-end billed cost. An arm that removed 38% of estimated raw tool-output tokens incurred 6.8% higher paired cost (95% CI: +2.8% to +11.3%), while per-task reduction showed only a weak association with cost change (Pearson r = 0.15). Third, aggressive compression can remove action-critical evidence: on SWE-bench-derived Go tasks, compression reduced successful patch application from 27/40 to 15/40 by corrupting verbatim edit anchors. We propose evaluating context-reduction systems by success-adjusted billed cost rather than token reduction alone.

Large Language Models (LLMs) have achieved remarkable success in complex reasoning tasks, but their inference remains computationally inefficient. We observe a common failure mode in many prevalent LLMs, overthinking, where models generate verbose and tangential reasoning traces even for simple queries. Recent works have tried to mitigate this by enforcing fixed token budgets, however, this can lead to underthinking, especially on harder problems. Through empirical analysis, we identify that this inefficiency often stems from unclear problem-solving strategies. To formalize this, we develop a theoretical model, BBAM (Bayesian Budget Allocation Model), which models reasoning as a sequence of sub-questions with varying uncertainty, and introduce the E^3 metric to capture the trade-off between correctness and computation efficiency. Building on theoretical results from BBAM, we propose Plan-and-Budget, a model-agnostic, test-time framework that decomposes complex queries into sub-questions and allocates token budgets based on estimated complexity using adaptive scheduling. Plan-and-Budget improves reasoning efficiency across a range of tasks and models, achieving up to +70% accuracy gains, -39% token reduction, and +187.5% improvement in E^3. Notably, it elevates a smaller model (DS-Qwen-32B) to match the efficiency of a larger model (DS-LLaMA-70B)-demonstrating Plan-and-Budget's ability to close performance gaps without retraining. Our code is available at anonymous.4open.science/r/P-and-B-6513/.

Multi-image vision-language models (VLMs) incur substantial computation when processing long visual sequences. Compressing these sequences must preserve complementary evidence across images and keep tokens from different images separate. Global selection can underrepresent question-relevant images, and pruning in early decoder layers risks removing relevant content before sufficient multimodal processing, yet existing methods spend most of the budget on these early layers. We propose SIEVE, a training-free, single-pass framework that jointly allocates visual tokens across images and decoder depth. SIEVE first pools each image into a compact spatial grid to reduce early processing cost. It then progressively selects representatives using prompt attention under equal per-image quotas and aggregates unselected tokens into similar representatives from the same image. These operations keep every image represented and its tokens separate from other images, while refining the retained representations. A closed-form schedule sets retention counts to target a prescribed average token budget across decoder layers. Across four VLM backbones, SIEVE achieves the highest average at a 6.25% budget on five multi-image benchmarks and four single-image benchmarks, and the highest mean on two video benchmarks. On LLaVA-OneVision-7B, it outperforms the strongest baseline by 9.96 and 9.03 points at 6.25% and 12.5% budgets, and at 12.5% it exceeds the uncompressed model by 2.01 points with 6.6x fewer prefill FLOPs.

Memory is increasingly central to Large Language Model (LLM) agents operating beyond a single context window, yet most existing systems rely on offline, query-agnostic memory construction that can be inefficient and may discard query-critical information. Although runtime memory utilization is a natural alternative, prior work often incurs substantial overhead and offers limited explicit control over the performance-cost trade-off. In this work, we present BudgetMem, a runtime agent memory framework for explicit, query-aware performance-cost control. BudgetMem structures memory processing as a set of memory modules, each offered in three budget tiers (i.e., Low/Mid/High). A lightweight router performs budget-tier routing across modules to balance task performance and memory construction cost, which is implemented as a compact neural policy trained with reinforcement learning. Using BudgetMem as a unified testbed, we study three complementary strategies for realizing budget tiers: implementation (method complexity), reasoning (inference behavior), and capacity (module model size). Across LoCoMo, LongMemEval, and HotpotQA, BudgetMem surpasses strong baselines when performance is prioritized (i.e., high-budget setting), and delivers better accuracy-cost frontiers under tighter budgets. Moreover, our analysis disentangles the strengths and weaknesses of different tiering strategies, clarifying when each axis delivers the most favorable trade-offs under varying budget regimes.

LLM-agent budget overruns are a documented production failure class: a single retry loop can spend thousands of dollars before an operator notices, and the in-process integrity properties that would prevent it (no aliasing, no double-spend, no use-after-delegation of a cost-bearing value) are enforced, if at all, by ad-hoc wrappers rather than by the type system. Our central contribution is empirical: a catalog of 63 confirmed production incidents from 21 orchestration frameworks (2023-2026), each backed by a quoted GitHub issue and, where reported, a dollar loss, organized into an eight-cluster failure taxonomy (inter-rater Cohen's kappa = 0.837, N = 113), plus 47 supplementary structural entries. As one mitigation evaluated against this taxonomy, we build token-budgets, an 1,180-line Rust crate (no unsafe) that operationalizes affine ownership so that cloning, double-spending, or using a budget after delegating it are compile errors rather than runtime hazards an operator must remember to avoid. The dollar cap is runtime arithmetic under an estimator assumption; the affine layer makes that arithmetic non-bypassable. On single-agent workloads a 4-line Python counter matches the crate at 0/30 overshoot, so the distinguishing value is non-bypassability under operator error in multi-agent delegation: the delegation-fanout race documented in 11 incidents is rejected by the borrow checker at compile time, while the same pattern under asyncio overshoots 30/30 and three disciplined alternatives overshoot 0/30. Across five runtimes, three providers, and a temperature-stratified live-API test (N = 160), the approach reports zero cap violations and zero false refusals, at operational parity with concurrent work. Static over-reservation is 4-6x (2.11x adaptive). Binary-level cap-soundness on the running binary is left open.

Software delegates more of its branches to models every year: which queue a ticket enters, whether a command is safe to run, whether a claim clears without a person. What the program needs back is not prose. It is one of n declared options and a number it can threshold. Today that costs a round trip to a frontier model -- hundreds of milliseconds, a per-token bill, and a parser -- for a question that is usually a conjunction of three clauses. this-that-model-1.0 is a 2B-parameter typed decision model. Its answer is read directly from the hidden state at a designated position and restricted to the option set the caller declared, so no text is generated, nothing can be malformed, and every question in a request is answered in the same forward pass. It decides in 30.9 ms on one laptop GPU and generates zero output tokens doing it, where a frontier API call costs 8758 ms and the hosted systems that answer these questions well spend between 21 and 212 generated tokens per question thinking first, billed for every one. It sustains 32 decisions per second on one consumer GPU and never lets the state leave the machine. On a third party's recorded cohort of 68 decision questions, on their inputs and their wording, it scores 0.941 with a Brier score of 0.042, against 0.765 and 0.133 for the hosted service Jev on the same items. One pass of our 42-family internal suite takes 32 seconds and 0.000217 USD of electricity; the most accurate hosted model we measured needs 155.2 minutes and 10.636 USD. We also report where it loses. On multi-step arithmetic, which a single forward pass cannot carry intermediate results through, it scores 0.560 against their 0.98 to 1.00, and a targeted second training round improved the five task families it was written for and transferred to none of the other 13. The model is open-sourced in https://huggingface.co/flock-io/this-that-model-1.0

When audio, video, and text disagree, accuracy alone does not show whether to acquire another source or abstain. We study these choices in a controlled three-score benchmark: a policy observes two signed scores, may request the third at a cost, and can abstain. The primary reward is mechanism-specific: abstention is correct only for one designated ambiguity mechanism and is penalized under mixed corruption. Matched controls show that a threshold policy matches always-request decisions with fewer requests; its advantage over always-answer fusion depends on the reward assigned to that ambiguity. On a partially held-out synthetic split, the threshold policy reaches 0.789 +/- 0.006 targeted decision accuracy and 0.481 +/- 0.014 utility across 83 seeds. A three-score majority reference reaches 0.626 +/- 0.008 and 0.252 +/- 0.016, but uses more information. In a matched-budget test, a train-only value selector improves utility over no-query and matched-random policies at 10% and 25% budgets, while pair uncertainty has higher utility at every budget. At 50% and 63.7% budgets, the selector lowers utility despite slightly higher non-ambiguous accuracy. If all abstentions are scored incorrect, majority outranks the threshold policy in utility. At a central temporal setting, full-trace controls match the neural models while position perturbations separate them. On held-out-actor emotion clips, eight-frame fusion has opposite-signed accuracy differences for two encoder pairs, with both actor intervals containing zero; matched-request routing gains are small and uncertain. These results separate full-modality accuracy from pre-request selection value and show that selection value depends on budget and the observed-pair ranking.

AI agents depend on repeated decisions to accomplish tasks, with large language models (LLMs) commonly serving as their core decision makers. The recent emergence of Jev offers an opportunity to make these decisions faster and at lower cost. However, it remains unclear whether such local savings can be translated into efficient and reliable task completion. In this paper, we investigate this efficiency transfer across five agents in two representative application settings, embodied tasks in ALFWorld and mobile app use in AndroidWorld. Across 4,500 task executions, we jointly evaluate task effectiveness, decision-level and task-level resource use, and execution behavior. Experimental results show that Jev integration substantially reduces decision cost and latency across all five agents, with task-level gains varying by workflow. Task success remains close to baseline levels on ALFWorld and declines more substantially in two mobile agents. Auxiliary reasoning and additional interactions absorb decision savings, while failed recovery and premature completion limit successful delivery. Mean cost per successful outcome increases in one agent even as spending per attempt falls. These findings highlight the importance of preserving progress, supporting recovery, and verifying completion when integrating efficient decision models into agent workflows. Their practical value should be assessed through successful task delivery, recovery behavior, and resources per delivered outcome.

Evaluation budgets in agentic retrieval-augmented generation span questions, search trajectories, and repeated answers. We measure allocation precision, reading efficiency, and cost boundaries using a retrieval-feedback comparison on HotpotQA and MuSiQue. At 34.14--34.39M model tokens, broader question coverage lowers standard error by 33\% versus five reads and 12.6\% versus three trajectories. Archived nested and Q-only forecasts predict these allocations within 4.0\% and 3.5\%, respectively. Depth subsets establish no clear forecasting advantage beyond the two-trajectory audit. One-read variance penalties relative to the fitted optimum at the same token budget are 0--9.9\%, with substantial Pro uncertainty. Under recorded model fees, more questions beat more trajectories at search prices of \$0--1 per 1,000 requests; question-versus-read fee rankings remain unresolved. Temperature zero cuts answer disagreement from 14.3\% to 3.4\% while comparison precision stays similar. \par\medskip\noindentKeywords: Agentic RAG; Evaluation budget; Generalizability theory; Repeated sampling.

As LLM agents evolve, tokens have emerged as the core economic primitives of Agentic AI. However, their exponential consumption introduces severe computational, collaborative, and security bottlenecks. Current surveys remain fragmented across system optimization, architecture design, and trust, lacking a unified framework to evaluate the fundamental trade-off between output quality and economic cost. To bridge this gap, this survey presents the first comprehensive survey of Token Economics. By unifying computer science and economics, we conceptualize tokens as production factors, exchange mediums, and units of account. We synthesize existing literature across a four-dimensional taxonomy: (1) Micro-level (Single Agent): Optimizing budget-constrained factor substitution via neoclassical firm theory. (2) Meso-level (Multi-Agent Systems): Minimizing collaboration friction using transaction cost and principal-agent theories. (3) Macro-level (Agent Ecosystems): Addressing congestion externalities and pricing via mechanism design. (4) Security: Internalizing adversarial threats as endogenous economic constraints. Finally, we outline frontier directions, including differentiable token budgets and dynamic markets, to lay the theoretical foundation for scalable next-generation agent systems.

Extractive prompt compression promises to cut LLM inference costs by removing low-information tokens, and learned compressors such as LLMLingua-2 report strong results on English benchmarks. Most other languages already pay a token premium: the same content costs 1.3-1.8x more tokens than in English. We ask whether compression closes or widens this gap. Using fully parallel data in ten languages spanning five scripts, with controls budget-matched in the target model's tokenizer, we audit four learned compressors against four deterministic baselines, on eleven target models from ten vendors (over 250,000 evaluation calls). Three of the compressors are trained with English supervision (LLMLingua-2 XLM-R/mBERT; Kompress-v2 from the production Headroom stack); the fourth, XProvence, is trained multilingually. First, the transfer gap is real, replicates across target models and compressor backbones, and is strongly rate-dependent: at a 0.33 keep-rate English retains 57-62% of normalized context utilization while Lithuanian retains 10-24% and Chinese essentially none, despite Chinese having the smallest token premium. Second, the gap tracks compression supervision data, not architecture. All three English-trained compressors show it, deterministic methods show no comparable gap, and the multilingually trained XProvence v1 shows none. Its v2 release, retrained on translated data, empties 92% of Chinese contexts at its aggressive threshold without any warning. Third, in a harder long-context setting, aggressive learned compression drives compressed contexts to or below no-context utility in three of five non-English languages. A translate-then-compress pipeline matches or beats native compression at roughly half the token cost in three of five tested languages. We release all code, compressions, and model outputs. Safe compression budgets are much smaller outside English.

High-resolution Vision-Language Models (VLMs) have been widely used in multimodal tasks to enhance accuracy by preserving detailed image information. However, these models often generate excessive visual tokens due to encoding multiple partitions of the input image. Processing these excessive visual tokens is computationally challenging, especially in resource-constrained environments with commodity GPUs. To support high-resolution images while meeting resource constraints, we propose High-Resolution Early Dropping (HiRED), a token-dropping scheme that operates within a fixed token budget before the Large Language Model (LLM) stage. HiRED can be integrated with existing high-resolution VLMs in a plug-and-play manner, as it requires no additional training while still maintaining superior accuracy. We strategically use the vision encoder's attention in the initial layers to assess the visual content of each image partition and allocate the token budget accordingly. Then, using the attention in the final layer, we select the most important visual tokens from each partition within the allocated budget, dropping the rest. Empirically, when applied to LLaVA-Next-7B on NVIDIA TESLA P40 GPU, HiRED with a 20% token budget increases token generation throughput by 4.7, reduces first-token generation latency by 15 seconds, and saves 2.3 GB of GPU memory for a single inference.

LLM inference is still evaluated mainly as a model or software problem: accuracy, latency, throughput, and hardware utilization. This is incomplete. At deployment scale, the relevant output is a quality-conditioned token produced under joint constraints from effective compute, delivered data-center power, cooling capacity, PUE, and utilization. We argue that the ML community should treat inference as energy-to-token production. We formalize this view with a dimensionally consistent Token Production Function in which token rate is bounded by both compute-per-token and energy-per-token ceilings. Listed API prices vary by over an order of magnitude across providers, but we use price dispersion only as directional motivation, not as causal evidence of marginal cost. The core physical question is instead: under fixed quality and service targets, when does the binding constraint move from theoretical peak compute toward delivered power, cooling, and operational efficiency? Under this framing, system optimizations -- latent KV-cache compression, sparse or heavily compressed attention, quantization, routing, and difficulty-adaptive reasoning -- are not merely local engineering tricks. They are energy-to-token levers because they reduce FLOPs/token, joules/token, memory traffic, or utilization losses under fixed (q^{*},s^{*}). We therefore call for inference papers and benchmarks to report Joules/token, active binding constraint, PUE-adjusted delivered power, and utilization-adjusted token output alongside accuracy and latency.

Long-horizon LLM inference turns the key--value (KV) cache into the dominant GPU memory consumer and makes per-token attention increasingly expensive. Many common eviction policies use static recency windows or historical attention, leaving unused a signal computed on every decoding step: the model's current uncertainty. We introduce CONF-KV, a KV-cache manager that converts the next-token distribution into a scalar confidence score and uses it to choose the per-step cache budget, retaining more context when the model is uncertain and pruning aggressively when it is confident. Within each budget, tokens are ranked by a composite of accumulated attention mass and recency, while a protected recent window preserves local coherence. We combine the policy with blockwise online-softmax attention, mixed FP16/INT8 storage, and a pyramidal per-layer budget variant. Across four model families and generated lengths up to 4K, CONF-KV stays near the footprint of a fixed 512-token sliding window while remaining within 1.5--2.1 perplexity points of full KV. On Needle-in-a-Haystack up to 32K tokens, CONF-KV reaches 91.4% retrieval accuracy versus 53.8% for sliding windows and 80.6% for H2O; on 75 VisualWebArena tasks it retains 95.3% of full-KV success at 2.8 times lower peak memory.

We investigate a failure mode of large language models (LLMs) in which plain-text prompts elicit excessive outputs, a phenomenon we term Overflow. Unlike jailbreaks or prompt injection, Overflow arises under ordinary interaction settings and can lead to elevated serving cost, latency, and cross-user performance degradation, particularly when scaled across many requests. Beyond usability, the stakes are economic and environmental: unnecessary tokens increase per-request cost and energy consumption, compounding into substantial operational spend and carbon footprint at scale. Moreover, Overflow represents a practical vector for compute amplification and service degradation in shared environments. We introduce BenchOverflow, a model-agnostic benchmark of nine plain-text prompting strategies that amplify output volume without adversarial suffixes or policy circumvention. Using a standardized protocol with a fixed budget of 5000 new tokens, we evaluate nine open- and closed-source models and observe pronounced rightward shifts and heavy tails in length distributions. Cap-saturation rates (CSR@1k/3k/5k) and empirical cumulative distribution functions (ECDFs) quantify tail risk; within-prompt variance and cross-model correlations show that Overflow is broadly reproducible yet heterogeneous across families and attack vectors. A lightweight mitigation-a fixed conciseness reminder-attenuates right tails and lowers CSR for all strategies across the majority of models. Our findings position length control as a measurable reliability, cost, and sustainability concern rather than a stylistic quirk. By enabling standardized comparison of length-control robustness across models, BenchOverflow provides a practical basis for selecting deployments that minimize resource waste and operating expense, and for evaluating defenses that curb compute amplification without eroding task performance.

PromptReps showed that an autoregressive language model can be used directly as a retriever by prompting it to generate dense and sparse representations of a query or passage. Extending this to multiple representatives is inefficient for autoregressive models, since tokens must be generated sequentially, and prior multi-token variants did not reliably improve over single-token decoding. We show that the bottleneck is sequential generation, not the multi-token idea itself. DiffRetriever is a representative-token retriever for diffusion language models: it appends K masked positions to the prompt and reads all K in a single bidirectional forward pass. Across in-domain and out-of-domain evaluation, multi-token DiffRetriever substantially improves over single-token on every diffusion backbone we test, while autoregressive multi-token is flat or negative and pays a latency cost that scales with K where diffusion does not. After supervised fine-tuning, DiffRetriever on Dream is the strongest BEIR-7 retriever in our comparison, ahead of PromptReps, the encoder-style DiffEmbed baseline on the same diffusion backbones, and the contrastively fine-tuned single-vector RepLLaMA. A per-query oracle on the frozen base model exceeds contrastive fine-tuning at the same fixed budget, pointing to adaptive budget selection as future work. Code is available at https://github.com/ielab/diffretriever.

Moving ML-mediated decision making onto privacy-preserving clients decentralises the economic decision along with the inference. Shared budget constraints then depend on information that cannot be globally current, creating an information-structure failure that conventional pacing is not designed to solve. We study this information misalignment in an auction-logic-faithful on-device simulation with 36 campaigns and 50 devices. Accounting is in dimensionless integer score units; no currency semantics are claimed. Across 30 paired demand paths, proportional Even pacing overspends 17.77% after one tick of staleness and 1,669.31% after 50 ticks under the original 20-times budget pressure. The effect does not depend on that severe a budget: at two-times pressure, 50-tick overspend remains 106.95%. A visible-budget no-sale guard makes zero-lag compliance exact at this score-unit granularity, yet leaves 11.88% overspend at one tick because other devices' debits remain invisible. A declared bursty, heterogeneous-device sweep retains a strictly increasing mean lag curve. We derive a finite-window expected excess-debit bound under conditional charge caps and find positive paired slack in every bounded-value cell. A second, incentive misalignment arises when the ML/pacing score transformation is allowed to change payment units: 98.23% of rival auctions at one tick admit a profitable deviation. An executable implementation-level counterexample isolates the runner-up's multiplier in the winner's price. Critical-base-bid payment is per-auction DSIC conditional on current multipliers, but does not establish dynamic truthfulness and does not repair base-value ranking disagreement.

Test-time scaling has emerged as an effective approach for improving language model performance by utilizing additional compute at inference time. Recent studies have shown that overriding end-of-thinking tokens (e.g., replacing "</think>" with "Wait") can extend reasoning steps and improve accuracy. In this work, we explore whether a dedicated continue-thinking token can be learned to trigger extended reasoning. We augment a distilled version of DeepSeek-R1 with a single learned "<|continue-thinking|>" token, training only its embedding via reinforcement learning while keeping the model weights frozen. Our experiments show that this learned token achieves improved accuracy on standard math benchmarks compared to both the baseline model and a test-time scaling approach that uses a fixed token (e.g., "Wait") for budget forcing. In particular, we observe that in cases where the fixed-token approach enhances the base model's accuracy, our method achieves a markedly greater improvement. For example, on the GSM8K benchmark, the fixed-token approach yields a 1.3% absolute improvement in accuracy, whereas our learned-token method achieves a 4.2% improvement over the base model that does not use budget forcing.

Hierarchical two-hop dispatch can reduce slow-fabric traffic in expert-parallel Mixture-of-Experts training, but it adds a second collective and an arrival-side operator chain. We present a cost model for screening that trade at the communication-call level, bounded by validation gates that withdraw a capability in code when they fail rather than reporting a caveat. At a reference geometry with 16 groups of 8 ranks, q=3, H=2048, and 4096 tokens per rank, the corrected effective breakeven hierarchy ratio is 3.98 for the measured PyTorch arrival chain, 1.49 for a hypothetical fused target, and 1.10 at zero implementation overhead. These are ratio-only sensitivity results, not deployment predictions: platform A measures 1.03, platform B has no separated fast/slow measurement, and neither machine measured here reaches the hierarchical regime. Four communication-level corpora pass their gates; a drift probe and the step-level gate fail. The latter failure is enforced in code, so we make no training-throughput prediction. The enabling routing constraint fixes per-token fan-out and per-selected-group quota, while aggregate per-peer counts remain data-dependent. Its measured validation-loss cost is small but nonzero (+0.0034 nats); downstream equivalence is reported with incomplete estimator provenance and is therefore not independently reconstructible from the artifact. Code, calibration constants and the validation gates are at https://github.com/weich97/TerraceMoE-simulator.

Multi-agent systems can improve reliability, yet under a fixed inference budget they often help, saturate, or even collapse. We develop a minimal and calibratable theory that predicts these regimes from three binding constraints of modern agent stacks: finite context windows, lossy inter-agent communication, and shared failures among similar agents. Each leaf agent is summarized by a compute-performance scaling exponent β; communication is captured by a message-length fidelity curve γ(m); dependence is captured by an effective shared-error correlation ρ; and a context window W imposes hard fan-in limits that make hierarchy necessary. For binary success/failure tasks with majority aggregation, we prove a sharp phase transition for deep b-ary trees with correlated inputs and lossy communication: a single scalar α_ρ (combining γ(m), ρ, and fan-in b) determines whether weak signal is amplified to a nontrivial fixed point or washed out to chance. In the amplifying regime, we derive an organization exponent s and show that budgeted synergy, i.e., outperforming the best single agent under the same total budget, occurs exactly when s>β, yielding closed-form compute allocation rules and explicit budget thresholds. We further characterize saturation via a mixing depth and provide a conservative clipped predictor that remains accurate across growth and saturation. A continuous-performance warm-up gives closed-form risks for star, chain, and tree organizations, making correlation- and communication-induced floors explicit and exposing the core design trade-offs in a smooth setting. Finally, we validate the predicted phase boundaries in controlled synthetic simulations and show how the same mechanisms explain the dominant bottlenecks reported in recent large-scale matched-budget studies of LLM agent-system scaling.

Increasing the thinking budget of AI models can significantly improve accuracy, but not all questions warrant the same amount of reasoning. Users may prefer to allocate different amounts of reasoning effort depending on how they value output quality versus latency and cost. To leverage this tradeoff effectively, users need fine-grained control over the amount of thinking used for a particular query, but few approaches enable such control. Existing methods require users to specify the absolute number of desired tokens, but this requires knowing the difficulty of the problem beforehand to appropriately set the token budget for a query. To address these issues, we propose Adaptive Effort Control, a self-adaptive reinforcement learning method that trains models to use a user-specified fraction of tokens relative to the current average chain-of-thought length for each query. This approach eliminates dataset- and phase-specific tuning while producing better cost-accuracy tradeoff curves compared to standard methods. Users can dynamically adjust the cost-accuracy trade-off through a continuous effort parameter specified at inference time. We observe that the model automatically learns to allocate resources proportionally to the task difficulty and, across model scales ranging from 1.5B to 32B parameters, our approach enables a 2-3x reduction in chain-of-thought length while maintaining or improving performance relative to the base model used for RL training.

Vision transformer has emerged as a new paradigm in computer vision, showing excellent performance while accompanied by expensive computational cost. Image token pruning is one of the main approaches for ViT compression, due to the facts that the complexity is quadratic with respect to the token number, and many tokens containing only background regions do not truly contribute to the final prediction. Existing works either rely on additional modules to score the importance of individual tokens, or implement a fixed ratio pruning strategy for different input instances. In this work, we propose an adaptive sparse token pruning framework with a minimal cost. Specifically, we firstly propose an inexpensive attention head importance weighted class attention scoring mechanism. Then, learnable parameters are inserted as thresholds to distinguish informative tokens from unimportant ones. By comparing token attention scores and thresholds, we can discard useless tokens hierarchically and thus accelerate inference. The learnable thresholds are optimized in budget-aware training to balance accuracy and complexity, performing the corresponding pruning configurations for different input instances. Extensive experiments demonstrate the effectiveness of our approach. Our method improves the throughput of DeiT-S by 50% and brings only 0.2% drop in top-1 accuracy, which achieves a better trade-off between accuracy and latency than the previous methods.

Coding-agent benchmarks evaluate whether a single uninterrupted agent can resolve a repository issue. Real software work is messier: tasks are interrupted, reassigned, reviewed, and resumed from partial states left by another agent or engineer. We study this missing dimension through handoff debt: the rediscovery cost imposed when a predecessor's work is opaque or incomplete. Our takeover protocol interrupts a coding agent at deterministic handoff points, freezes the repository, and evaluates successor agents under four handoff views: repository state only, raw trace, summary notes, and structured notes. Across 75 source tasks, the protocol generates 181 handoff-point tasks and 724 takeover runs per successor model. Across three successor models, context-bearing handoffs reduce median agent events by 20-59% and cumulative prompt tokens by 42-63% relative to repository-only takeover. Solved-rate effects are smaller and model-dependent, but efficiency gains are consistent. These findings suggest that coding-agent evaluation should report not only whether a task is solved, but also how costly that work is for another agent to resume.

While vision-language-action (VLA) models have advanced rapidly, running them at every control step remains expensive. Prior work accelerates VLA inference using techniques like action chunking and visual-token pruning, typically evaluating based on latency and average task success. However, acceleration may discard information and break tasks the original policy would solve, a risk hidden by average metrics. Measuring these failures is challenging because action deviations compound over closed-loop trajectories, meaning task failure is only observable across full episodes. We therefore define an acceleration-induced failure via paired rollouts from identical initial conditions, tracking when the reference succeeds but the accelerated policy fails. To manage this, we introduce CARE, an approach for certified accelerator selection. CARE uses paired rollouts on a calibration set to provide finite-sample guarantees that acceleration-induced failure risk stays below a user-specified budget. It deploys the fastest certified candidate, falling back to the reference if none qualify. By relying only on terminal outcomes and measured compute, CARE applies unchanged across diverse acceleration mechanisms, while sequential testing and failure-triggered reference rollouts keep certification affordable. On four LIBERO suites with OpenVLA-OFT, CARE certifies 9.0--10.8times speedups while guaranteeing (at 95% confidence) that at least 85.8% of reference-solved episodes are preserved. Under tight budgets, selectors without guarantees exceed the budget in up to 75% of trials, whereas CARE stays within budget and its sequential form uses 78.9% fewer rollouts than exhaustive evaluation. CARE further generalizes to flow-step reduction for π_{0.5}, and to Qwen3.5-9B and Llama-3.1-8B agents in Crafter.

Token-choice Mixture-of-Experts (TC-MoE) routes each token to a fixed number of experts, limiting dynamic computation allocation and requiring auxiliary losses to maintain load balance. We propose Expert Threshold (ET) routing, where each expert maintains an exponential moving average (EMA) threshold estimated from the global token distribution. At both training and inference, each token is independently routed to an expert if its score exceeds the expert's threshold, enabling dynamic computation allocation while achieving load balance without auxiliary losses. This fully causal mechanism eliminates dependence on other tokens in the batch, making it well-suited for autoregressive language modeling. In pretraining experiments scaling to 2.4B parameters on FineWeb-Edu, ET achieves 0.067 lower cross-entropy loss than TC-MoE, equivalent to reaching the same performance with 1.6times fewer tokens.

Intent-based Open RAN needs an interpreter that turns intents into A1 policies within the loop of the RAN intelligent controller (RIC). Decision models such as Jev-1.13.0 return typed policy fields, whereas generative large language models (LLMs) produce the policy token by token. We ask whether the extra delay of LLMs costs control deadlines, RIC capacity, or radio performance. We compare Jev-1.13.0 and two other decision models with LLMs on the RANIntent v1 benchmark, in closed-loop ns-3 simulation and on a real A1 and E2 path. Median interpretation takes 0.286 to 2.35 s, against under 25 ms for A1 and E2 transfer. Jev-1.13.0 meets the 1 s near-real-time budget on 99.8% of calls, while two hosted LLMs meet it on 17.9% and 0%. In the radio network, ideal enforcement moves the affected-class service-level agreement (SLA) violation by 3.96 percentage points in the direction each intent requests, against no update at the base point. No hosted LLM showed a resolved increase over Jev-1.13.0 at that point. At the same point, per-second direct control gave no resolved SLA reduction over a numerical xApp. Slow interpreters miss the 1 s budget, and two interpreters saturate their queues at 2 intents/s, whereas no radio penalty of slow interpreters was resolved at the base point.

Large language models (LLMs) can achieve strong reasoning performance with sufficient computation, but they do not inherently know how much computation a task requires. We study budgeted inference-time reasoning for multiple tasks under a strict global token constraint and formalize it as a Ordered Stochastic Multiple-Choice Knapsack Problem(OS-MCKP). This perspective highlights a meta-cognitive requirement -- anticipating task difficulty, estimating return over investment (ROI), and allocating computation strategically. We propose ROI-Reasoning, a two-stage framework that endows LLMs with intrinsic, budget-aware rationality. In the first stage, Meta-Cognitive Fine-Tuning teaches models to predict reasoning cost and expected utility before generation, enabling explicit solve-or-skip decisions. Next, Rationality-Aware Reinforcement Learning optimizes sequential decision making under a hard token budget, allowing models to learn long-horizon allocation strategies. Across budgeted mathematical reasoning benchmarks, ROI-Reasoning consistently improves overall score while substantially reducing regret under tight computation budgets.

Scaling test-time compute is crucial for enhancing the reasoning capabilities of large language models (LLMs). Existing approaches typically employ reinforcement learning (RL) to maximize a verifiable reward obtained at the end of reasoning traces. However, such methods optimize only the final performance under a large and fixed token budget, which hinders efficiency in both training and deployment. In this work, we present a novel framework, AnytimeReasoner, to optimize anytime reasoning performance, which aims to improve token efficiency and the flexibility of reasoning under varying token budget constraints. To achieve this, we truncate the complete thinking process to fit within sampled token budgets from a prior distribution, compelling the model to summarize the optimal answer for each truncated thinking for verification. This introduces verifiable dense rewards into the reasoning process, facilitating more effective credit assignment in RL optimization. We then optimize the thinking and summary policies in a decoupled manner to maximize the cumulative reward. Additionally, we introduce a novel variance reduction technique, Budget Relative Policy Optimization (BRPO), to enhance the robustness and efficiency of the learning process when reinforcing the thinking policy. Empirical results in mathematical reasoning tasks demonstrate that our method consistently outperforms GRPO across all thinking budgets under various prior distributions, enhancing both training and token efficiency.

Recent work shows that, beyond discrete reasoning through explicit chain-of-thought steps, which are limited by the boundaries of natural languages, large language models (LLMs) can also reason continuously in latent space, allowing richer information per step and thereby improving token efficiency. Despite this promise, latent reasoning still faces two challenges, especially in training-free settings: 1) purely latent reasoning broadens the search distribution by maintaining multiple implicit paths, which diffuses probability mass, introduces noise, and impedes convergence to a single high-confidence solution, thereby hurting accuracy; and 2) overthinking persists even without explicit text, wasting tokens and degrading efficiency. To address these issues, we introduce SwiReasoning, a training-free framework for LLM reasoning which features two key innovations: 1) SwiReasoning dynamically switches between explicit and latent reasoning, guided by block-wise confidence estimated from entropy trends in next-token distributions, to balance exploration and exploitation and promote timely convergence. 2) By limiting the maximum number of thinking-block switches, SwiReasoning curbs overthinking and improves token efficiency across varying problem difficulties. On widely used mathematics and STEM benchmarks, SwiReasoning consistently improves average accuracy by 1.5%-2.8% across reasoning LLMs of different model families and scales. Furthermore, under constrained budgets, SwiReasoning improves average token efficiency by 56%-79%, with larger gains as budgets tighten.

Recent work on enhancing the reasoning abilities of large language models (LLMs) has introduced explicit length control as a means of constraining computational cost while preserving accuracy. However, existing approaches rely on fixed-length training budgets, which do not take advantage of the natural progression from exploration to compression during learning. In this work, we propose a curriculum learning strategy for length-controlled reasoning using Group Relative Policy Optimization (GRPO). Our method starts with generous token budgets and gradually tightens them over training, encouraging models to first discover effective solution strategies and then distill them into more concise reasoning traces. We augment GRPO with a reward function that balances three signals: task correctness (via verifier feedback), length efficiency, and formatting adherence (via structural tags). Experiments on GSM8K, MATH500, SVAMP, College Math, and GSM+ demonstrate that curriculum-based training consistently outperforms fixed-budget baselines at the same final budget, achieving higher accuracy and significantly improved token efficiency. We further ablate the impact of reward weighting and decay schedule design, showing that progressive constraint serves as a powerful inductive bias for training efficient reasoning models. Our code and checkpoints are released at: https://github.com/hammoudhasan/curriculum_grpo.

Providers of LLM-as-a-service have predominantly adopted a simple pricing model: users pay a fixed price per token. Consequently, one may think that the price two different users would pay for the same output string under the same input prompt is the same. In our work, we show that, surprisingly, this is not (always) true. We find empirical evidence that, particularly for non-english outputs, both proprietary and open-weights LLMs often generate the same (output) string with multiple different tokenizations, even under the same input prompt, and this in turn leads to arbitrary price variation. To address the problem of tokenization multiplicity, we introduce canonical generation, a type of constrained generation that restricts LLMs to only generate canonical tokenizations -- the unique tokenization in which each string is tokenized during the training process of an LLM. Further, we introduce an efficient sampling algorithm for canonical generation based on the Gumbel-Max trick. Experiments on a variety of natural language tasks demonstrate that our sampling algorithm for canonical generation is comparable to standard sampling in terms of performance and runtime, and it solves the problem of tokenization multiplicity.

Token-pruning policies are usually designed for a single recognition pipeline, but pretrained Vision Transformers are reused across tasks with different spatial demands. We ask which parts of a pruning policy transfer across image classification, semantic segmentation, and object detection. For each pipeline, controlled probes freeze the no-pruning checkpoint and apply a series of parameter-free reduction criteria at one eligible layer at a time without retraining. The probes reveal three differences: segmentation and detection rank the criteria differently, classification is especially sensitive to attention-based pruning in the earliest layers, and the dense tasks prefer opposite recovery endpoints. These findings motivate Task-Adaptive Pruning (TAP). Existing register tokens serve as task-agnostic storage for feature artifacts. TAP instead introduces one task register per task and activates only the current one. Its evolving state ranks tokens, distributes an exact removal budget over depth, and sets the recovery scale for dense features. At a final keep rate of ρ=0.5, our jointly adapted model, TAP-J, reaches 47.0 mIoU at 1.30times encoder throughput on ADE20K and 53.7 box AP at 1.32times encoder throughput on COCO while remaining competitive on ImageNet-1K.

This position paper argues that agentic AI systems should be designed and evaluated as marginal token allocation economies rather than as text generators priced by the unit. We follow a single request -- a developer asking a coding agent to fix a failing test -- through four economic layers that today are designed in isolation: a router that decides which model answers, an agent that decides whether to plan, act, verify, or defer, a serving stack that decides how to produce each token, and a training pipeline that decides whether the trace is worth learning from. We show that all four layers are solving the same first-order condition -- marginal benefit equals marginal cost plus latency cost plus risk cost -- with different index sets and different prices. The framing is deliberately minimal: we do not propose a complete theory of AI economics. But adopting marginal token allocation as the shared accounting object explains why systems that locally minimize tokens globally misallocate them, predicts a small set of recurring failure modes (over-routing, over-delegation, under-verification, serving congestion, stale rollouts, cache misuse), and points to a concrete research agenda in token-aware evaluation, autonomy pricing, congestion-priced serving, and risk-adjusted RL budgeting.

LLM agents performing long-horizon tasks accumulate tool results that later steps may need. Passing the full history to every invocation is costly even when it fits within the context window, while reducing it risks omitting needed information. Existing context management methods can overlook how earlier tool results are used in subsequent execution, leaving needed information out of context. We introduce ContextRender, which manages context through a persistent graph of execution dependencies. We develop Tool-Flow Analysis to track how later operations reuse information from earlier tool results, providing a signal called observed reuse. A renderer combines this signal with recency and semantic relevance to select results within a fixed history budget, retaining omitted results for later use. Across AppWorld and 8-objective QA with three execution models, ContextRender outperforms the evaluated context management baselines using a 6K history budget, well below the models' maximum context windows. Within this budget, it achieves task performance close to or above that of passing the full history while reducing mean inference cost by 10.2%-32.2% relative to Full history. Ablations show that observed reuse improves task performance and retention of results reused later.

Advances in frontier coding agents have sharply increased the token costs of AI-assisted software engineering. Their token efficiency is typically evaluated on isolated tasks, whereas real-world developers use them to continuously build, update, and maintain large-scale projects. This mismatch motivates us to evaluate agents over sequences of realistic coding tasks, where we find that agents exhibit recurring environment-dependent token costs that existing methods struggle to reduce. To minimize these costs, we introduce TokenEconomizer, an external controller that cuts environment-specific token costs over a sequence of tasks. These cuts come from determining (i) which operational micro-skills to inject, (ii) which persistently irrelevant tools to prune, and (iii) when to intervene in unproductive failure loops. Across Claude Code, Codex, and GitHub Copilot CLI paired with frontier models, TokenEconomizer reduces the average token cost by 18.4-30.9% with comparable task success. Savings hold in every repository under every tested agent-model setting and generalize to diverse domains such as sequential web development and security auditing, where costs fall by 6.4-21.2% depending on task structure. Our work establishes adaptive black-box cost control as a promising direction for reducing token costs of coding agents in realistic settings.

The economic value of inference depends on how capacity and task information are distributed across stages of AI production. We study these organizational margins using controlled workflow experiments on externally verified software-engineering tasks. In two matched resource panels, direct execution records the same success rate of 59.6 percent at logical-token ceilings of 12,000 and 24,000, while success under information-constrained planning rises from 36.2 to 51.2 percent. The planning disadvantage narrows by 15.0 percentage points (95 percent task-cluster bootstrap interval: 4.2 to 25.8). A strict read-only planning campaign varies whether the planner sees the task issue. At 12,000 tokens, issue access raises success by about 16 percentage points over issue-hidden planning. Compared with direct execution, task-informed planning is about 10 points lower at 12,000 tokens; at 24,000 tokens, it shows a 29.6-point advantage. In the resource panels, direct execution uses substantially less than either ceiling, while the planning workflow's binding rate falls from 46.2 to 0.8 percent and downstream execution accounts for 89.9 percent of the increase in total use. Scale determines the capacity available to a system; workflow and information structure shape the productive value

Language models have graduated from being research prototypes to commercialized products offered as web APIs, and recent works have highlighted the multilingual capabilities of these products. The API vendors charge their users based on usage, more specifically on the number of ``tokens'' processed or generated by the underlying language models. What constitutes a token, however, is training data and model dependent with a large variance in the number of tokens required to convey the same information in different languages. In this work, we analyze the effect of this non-uniformity on the fairness of an API's pricing policy across languages. We conduct a systematic analysis of the cost and utility of OpenAI's language model API on multilingual benchmarks in 22 typologically diverse languages. We show evidence that speakers of a large number of the supported languages are overcharged while obtaining poorer results. These speakers tend to also come from regions where the APIs are less affordable to begin with. Through these analyses, we aim to increase transparency around language model APIs' pricing policies and encourage the vendors to make them more equitable.

Fine tuning has been regarded as a de facto approach for adapting large language models (LLMs) to downstream tasks, but the high training memory consumption inherited from LLMs makes this process inefficient. Among existing memory efficient approaches, activation-related optimization has proven particularly effective, as activations consistently dominate overall memory consumption. Although prior arts offer various activation optimization strategies, their data-agnostic nature ultimately results in ineffective and unstable fine tuning. In this paper, we propose TokenSeek, a universal plugin solution for various transformer-based models through instance-aware token seeking and ditching, achieving significant fine-tuning memory savings (e.g., requiring only 14.8% of the memory on Llama3.2 1B) with on-par or even better performance. Furthermore, our interpretable token seeking process reveals the underlying reasons for its effectiveness, offering valuable insights for future research on token efficiency. Homepage: https://runjia.tech/iclr_tokenseek/

Transformer-based language models spread FLOPs uniformly across input sequences. In this work we demonstrate that transformers can instead learn to dynamically allocate FLOPs (or compute) to specific positions in a sequence, optimising the allocation along the sequence for different layers across the model depth. Our method enforces a total compute budget by capping the number of tokens (k) that can participate in the self-attention and MLP computations at a given layer. The tokens to be processed are determined by the network using a top-k routing mechanism. Since k is defined a priori, this simple procedure uses a static computation graph with known tensor sizes, unlike other conditional computation techniques. Nevertheless, since the identities of the k tokens are fluid, this method can expend FLOPs non-uniformly across the time and model depth dimensions. Thus, compute expenditure is entirely predictable in sum total, but dynamic and context-sensitive at the token-level. Not only do models trained in this way learn to dynamically allocate compute, they do so efficiently. These models match baseline performance for equivalent FLOPS and wall-clock times to train, but require a fraction of the FLOPs per forward pass, and can be upwards of 50\% faster to step during post-training sampling.

We study reliability in autonomous language-model agents that translate user mandates into validated tool actions under real capital. The setting is DX Terminal Pro, a 21-day deployment in which 3,505 user-funded agents traded real ETH in a bounded onchain market. Users configured vaults through structured controls and natural-language strategies, but only agents could choose normal buy/sell trades. The system produced 7.5M agent invocations, roughly 300K onchain actions, about $20M in volume, more than 5,000 ETH deployed, roughly 70B inference tokens, and 99.9% settlement success for policy-valid submitted transactions. Long-running agents accumulated thousands of sequential decisions, including 6,000+ prompt-state-action cycles for continuously active agents, yielding a large-scale trace from user mandate to rendered prompt, reasoning, validation, portfolio state, and settlement. Reliability did not come from the base model alone; it emerged from the operating layer around the model: prompt compilation, typed controls, policy validation, execution guards, memory design, and trace-level observability. Pre-launch testing exposed failures that text-only benchmarks rarely measure, including fabricated trading rules, fee paralysis, numeric anchoring, cadence trading, and misread tokenomics. Targeted harness changes reduced fabricated sell rules from 57% to 3%, reduced fee-led observations from 32.5% to below 10%, and increased capital deployment from 42.9% to 78.0% in an affected test population. We show that capital-managing agents should be evaluated across the full path from user mandate to prompt, validated action, and settlement.

Test-time scaling is a promising new approach to language modeling that uses extra test-time compute to improve performance. Recently, OpenAI's o1 model showed this capability but did not publicly share its methodology, leading to many replication efforts. We seek the simplest approach to achieve test-time scaling and strong reasoning performance. First, we curate a small dataset s1K of 1,000 questions paired with reasoning traces relying on three criteria we validate through ablations: difficulty, diversity, and quality. Second, we develop budget forcing to control test-time compute by forcefully terminating the model's thinking process or lengthening it by appending "Wait" multiple times to the model's generation when it tries to end. This can lead the model to double-check its answer, often fixing incorrect reasoning steps. After supervised finetuning the Qwen2.5-32B-Instruct language model on s1K and equipping it with budget forcing, our model s1 exceeds o1-preview on competition math questions by up to 27% (MATH and AIME24). Further, scaling s1 with budget forcing allows extrapolating beyond its performance without test-time intervention: from 50% to 57% on AIME24. Our model, data, and code are open-source at https://github.com/simplescaling/s1.

MCP standardizes how LLMs interact with external systems, forming the foundation for general agents. However, existing MCP benchmarks remain narrow in scope: they focus on read-heavy tasks or tasks with limited interaction depth, and fail to capture the complexity and realism of real-world workflows. To address this gap, we propose MCPMark, a benchmark designed to evaluate MCP use in a more realistic and comprehensive manner. It consists of 127 high-quality tasks collaboratively created by domain experts and AI agents. Each task begins with a curated initial state and includes a programmatic script for automatic verification. These tasks demand richer and more diverse interactions with the environment, involving a broad range of create, read, update, and delete (CRUD) operations. We conduct a comprehensive evaluation of cutting-edge LLMs using a minimal agent framework that operates in a tool-calling loop. Empirical results show that the best-performing model, gpt-5-medium, reaches only 52.56\% pass@1 and 33.86\% pass^4, while other widely regarded strong models, including claude-sonnet-4 and o3, fall below 30\% pass@1 and 15\% pass^4. On average, LLMs require 16.2 execution turns and 17.4 tool calls per task, significantly surpassing those in previous MCP benchmarks and highlighting the stress-testing nature of MCPMark.

Diffusion language models can write several tokens in one forward pass, and they decide during generation how many passes to spend. Two outputs of the same token length can therefore vary a lot in cost. To understand the underlying factors of this cost variation, we analyze two diffusion models across 10 tasks. We find that the number of forward passes per token depends heavily on the task. Text copied or lightly edited from the input is cheap, while open-ended writing is costly. In DiffusionGemma, generating a token for story writing requires more than three times as many forward passes as for grammar correction. Two output properties explain much of this gap: how much of it is new relative to the input, and how rare its words are within the task. The same holds within a single response when thinking is involved. The final answer is cheap when it repeats the model's reasoning trace and costly when it adds something new. In multilingual generation, we find that highly tokenized languages also require more passes per token, adding to the cost they already pay with more tokens. Finally, cost tracks correctness: on the same math problem, wrong answers take more passes per token than correct ones, and weighting candidate answers by decoding speed beats majority voting by up to 4.2 points at no extra cost. Taken together, our results show that diffusion's computational advantage depends on how much of the output is already determined, while its decoding effort offers a practical signal for making inference more efficient and reliable.

Visual-token pruning can substantially reduce the inference cost of multimodal large language models (MLLMs), yet existing methods largely rely on fixed, handcrafted heuristics and costly expert trial and error. As pruning objectives, budgets, and model architectures diversify, manually navigating the expanding design space becomes increasingly difficult. This paper aims to build an AI4AI framework for visual-token pruning by addressing a natural question: Can large language models automatically design effective visual-token reduction algorithms? Although LLMs possess broad algorithmic knowledge and strong reasoning capabilities, translating such general knowledge into effective solutions for a specialized task remains nontrivial. We argue that the key lies in designing an appropriate search-state representation that connects the internal knowledge of LLMs with the structural requirements and constraints of visual-token pruning. Based on this insight, we propose AutoPrune, a training-free framework for LLM-driven visual-token pruning policy design. At its core, AutoPrune introduces a Token Pruning Domain-Specific Language (TPDSL) comprising 131 reusable atoms for budget control, token scoring, selection constraints, and token reassembly. A key property of TPDSL is that it represents each search state as a residual modification of a strong base policy. This residual formulation narrows the search space and directs the LLM's attention toward the policy components that are most consequential for performance. Experiments on 14 multimodal benchmarks and three MLLM backbones demonstrate the effectiveness, efficiency, and transferability of AutoPrune. Even when removing 94.4% of visual tokens, AutoPrune preserves more than 99% of full-token performance while reducing FLOPs by 9.9x and prefill latency by 6.4x.

Frontier LLMs are increasingly utilised across academia, society and industry. A commonly used unit for comparing models, their inputs and outputs, and estimating inference pricing is the token. In general, tokens are used as a stable currency, assumed to be broadly consistent across tokenizers and contexts, enabling direct comparisons. However, tokenization varies significantly across models and domains of text, making naive interpretation of token counts problematic. We quantify this variation by providing a comprehensive empirical analysis of tokenization, exploring the compression of sequences to tokens across different distributions of textual data. Our analysis challenges commonly held heuristics about token lengths, finding them to be overly simplistic. We hope the insights of our study add clarity and intuition toward tokenization in contemporary LLMs.

Recently, the scale of transformers has grown rapidly, which introduces considerable challenges in terms of training overhead and inference efficiency in the scope of task adaptation. Existing works, namely Parameter-Efficient Fine-Tuning (PEFT) and model compression, have separately investigated the challenges. However, PEFT cannot guarantee the inference efficiency of the original backbone, especially for large-scale models. Model compression requires significant training costs for structure searching and re-training. Consequently, a simple combination of them cannot guarantee accomplishing both training efficiency and inference efficiency with minimal costs. In this paper, we propose a novel Parallel Yielding Re-Activation (PYRA) method for such a challenge of training-inference efficient task adaptation. PYRA first utilizes parallel yielding adaptive weights to comprehensively perceive the data distribution in downstream tasks. A re-activation strategy for token modulation is then applied for tokens to be merged, leading to calibrated token features. Extensive experiments demonstrate that PYRA outperforms all competing methods under both low compression rate and high compression rate, demonstrating its effectiveness and superiority in maintaining both training efficiency and inference efficiency for large-scale foundation models. Our code will be released to the public.

Agent skills are often injected in full on every request, increasing token cost. We compare four content-preserving loading methods: Full, Skill Block, Reference, and Hybrid. Across SearchQA, SpreadsheetBench, ALFWorld, ScienceWorld, and SynthProc, we measure token usage using raw input for single-turn tasks and cache-correct effective input for multi-turn tasks. Results show no universal winner. Hybrid reduces input by 27.4% on SearchQA and 39.8% on SpreadsheetBench. On large multi-turn skills, Skill Block and Hybrid achieve substantial reductions, reaching 62.5% and 52.8% on ScienceWorld and 73.0% and 66.6% on SynthProc. ALFWorld shows smaller gains because procedures are short and repeatedly needed. Paired outcome tests detect no quality differences, though they do not establish equivalence. Overall, conditional loading is most beneficial when large portions of a skill are not needed on every turn.

Autoregressive (AR) language models generate text one token at a time, even when consecutive tokens are highly predictable given earlier context. We introduce MARS (Mask AutoRegreSsion), a lightweight fine-tuning method that teaches an instruction-tuned AR model to predict multiple tokens per forward pass. MARS adds no architectural modifications, no extra parameters, and produces a single model that can still be called exactly like the original AR model with no performance degradation. Unlike speculative decoding, which maintains a separate draft model alongside the target, or multi-head approaches such as Medusa, which attach additional prediction heads, MARS requires only continued training on existing instruction data. When generating one token per forward pass, MARS matches or exceeds the AR baseline on six standard benchmarks. When allowed to accept multiple tokens per step, it maintains baseline-level accuracy while achieving 1.5-1.7x throughput. We further develop a block-level KV caching strategy for batch inference, achieving up to 1.71x wall-clock speedup over AR with KV cache on Qwen2.5-7B. Finally, MARS supports real-time speed adjustment via confidence thresholding: under high request load, the serving system can increase throughput on the fly without swapping models or restarting, providing a practical latency-quality knob for deployment.

Test-time scaling can improve large language model reasoning by generating and combining multiple candidate responses. In sampling-based methods, the inference budget is often described by the number of generated candidates, N. However, N tells us how many candidates are generated, not how they are executed. The same candidate budget can be produced in one batched generation call or split across several sequential calls with smaller batch sizes. We first study the effect of increasing N on reasoning accuracy using Phi-3-mini and Qwen2.5-1.5B on 500 GSM8K prompts. As expected, increasing N from 1 to 8 improves accuracy by 8.4 percentage points for Phi-3-mini and 18.4 points for Qwen2.5-1.5B. However, accuracy alone does not show the systems cost of using a larger candidate budget. We therefore fix N = 8 and compare four generation schedules: 1x8, 2x4, 4x2, and 8x1, where axb denotes a generation calls with b candidates per call. We measure latency, throughput, GPU-hours, and gross GPU-device energy while keeping the total candidate count fixed. On A100 GPUs, eight serial calls use 4.64-4.86x as much gross GPU-device energy and have 5.77-6.12x the P95 latency of one batched call with eight candidates. The same pattern appears across three independently scheduled A100 nodes per model and in short-output SciQ/V100 experiments. These results show that candidate count alone is not enough to describe the systems cost of multi-candidate test-time scaling. When candidates are independent and memory allows it, fewer generation calls with larger batch sizes are more efficient. Evaluations should therefore report not only candidate count and accuracy, but also generation schedule and GPU-level systems metrics.

Test-time reasoning is increasingly used as a serving-time control knob, but extra reasoning is not uniformly valuable: it can repair failed attempts, waste compute on already-correct answers, or introduce harmful answer changes. We study this as a deployment allocation problem rather than a new-verifier problem. We introduce \sevra, Selective Verification for Reasoning Allocation, a serving-layer controller that decides whether to preserve a frozen solver's initial answer or invoke active verification. Using a frozen Qwen3-4B solver, we log intervention outcomes and train recoverability-aware gates from serving-visible attempt state. On \mathfive, selective verification reaches 76.3\% accuracy, compared with 75.5\% for always verifying, while reducing post-generation tokens by 26.8\% and harmful flips from 2.2\% to 1.0\%. However, an 8,192-token initial solve reaches 76.0\% accuracy with 28\% fewer total model tokens, showing that selective recovery is useful but not the best tested cost frontier. In frozen transfer to \gsm, the selective policy verifies only 3.0\% of examples, improves accuracy from 93.4\% to 94.5\%, and reduces verification tokens by 91.2\% relative to always verifying; again, a longer initial solve matches its accuracy with fewer realized tokens. On CommonsenseQA, always-on verification hurts, while Self-Consistency@5 improves accuracy at about five times the realized token cost. The resulting deployment rule is: tune the initial budget first, then use selective recovery when explicit checks, bounded retries, auditability, or regression-risk control matter.

Multimodal Large Language Models (MLLMs) suffer from substantial computational overhead due to the high redundancy in visual token sequences. Existing approaches typically address this issue using single-layer Vision Transformer (ViT) features and static pruning strategies. However, such fixed configurations are often brittle under diverse instructions. To overcome these limitations, we propose CLASP, a plug-and-play token reduction framework based on class-adaptive layer fusion and dual-stage pruning. Specifically, CLASP first constructs category-specific visual representations through multi-layer vision feature fusion. It then performs dual-stage pruning, allocating the token budget between attention-salient pivot tokens for relevance and redundancy-aware completion tokens for coverage. Through class-adaptive pruning, CLASP enables prompt-conditioned feature fusion and budget allocation, allowing aggressive yet robust visual token reduction. Extensive experiments demonstrate that CLASP consistently outperforms existing methods across a wide range of benchmarks, pruning ratios, and MLLM architectures. Code will be available at https://github.com/Yunkaidang/CLASP.

This paper studies close-loop task planning, which refers to the process of generating a sequence of skills (a plan) to accomplish a specific goal while adapting the plan based on real-time observations. Recently, prompting Large Language Models (LLMs) to generate actions iteratively has become a prevalent paradigm due to its superior performance and user-friendliness. However, this paradigm is plagued by two inefficiencies: high token consumption and redundant error correction, both of which hinder its scalability for large-scale testing and applications. To address these issues, we propose Tree-Planner, which reframes task planning with LLMs into three distinct phases: plan sampling, action tree construction, and grounded deciding. Tree-Planner starts by using an LLM to sample a set of potential plans before execution, followed by the aggregation of them to form an action tree. Finally, the LLM performs a top-down decision-making process on the tree, taking into account real-time environmental information. Experiments show that Tree-Planner achieves state-of-the-art performance while maintaining high efficiency. By decomposing LLM queries into a single plan-sampling call and multiple grounded-deciding calls, a considerable part of the prompt are less likely to be repeatedly consumed. As a result, token consumption is reduced by 92.2% compared to the previously best-performing model. Additionally, by enabling backtracking on the action tree as needed, the correction process becomes more flexible, leading to a 40.5% decrease in error corrections. Project page: https://tree-planner.github.io/

Large language models serve applications with diverse inference budgets, but training and maintaining separate models for each budget incurs substantial costs. Recent Looped Transformers enable flexible inference computation through parameter reuse, but maintaining strong performance across different inference budgets remains an open challenge. We introduce Accordion, a mixture-of-experts-based model that supports five inference budgets through fixed-prefix layer looping. Each budget follows a pre-defined execution path, while the minimum budget retains the unlooped model. Joint pretraining samples one budget per micro-batch and optimizes its next-token prediction loss without additional cross-budget consistency objectives. We further preserve these inference budgets through supervised fine-tuning and reinforcement learning, where parameter-group merging and budget-matched multi-teacher on-policy distillation integrate three domain experts into one model. After 3.6T tokens of pretraining, increasing inference depth improves Accordion-6B-A0.5B's five-category macro-average across 37 tasks by 3.01 percentage points, with the middle budget capturing 89% of this gain using half as many additional layer executions. The final model surpasses each domain expert on its corresponding task-and-budget average, while achieving strong instruction-following performance across all budgets and retaining depth-dependent gains in mathematical reasoning and code generation.

We utilize an offline reinforcement learning (RL) model for sequential targeted promotion in the presence of budget constraints in a real-world business environment. In our application, the mobile app aims to boost customer retention by sending cash bonuses to customers and control the costs of such cash bonuses during each time period. To achieve the multi-task goal, we propose the Budget Constrained Reinforcement Learning for Sequential Promotion (BCRLSP) framework to determine the value of cash bonuses to be sent to users. We first find out the target policy and the associated Q-values that maximizes the user retention rate using an RL model. A linear programming (LP) model is then added to satisfy the constraints of promotion costs. We solve the LP problem by maximizing the Q-values of actions learned from the RL model given the budget constraints. During deployment, we combine the offline RL model with the LP model to generate a robust policy under the budget constraints. Using both online and offline experiments, we demonstrate the efficacy of our approach by showing that BCRLSP achieves a higher long-term customer retention rate and a lower cost than various baselines. Taking advantage of the near real-time cost control method, the proposed framework can easily adapt to data with a noisy behavioral policy and/or meet flexible budget constraints.

Autoregressive large language model inference repeatedly invokes the target model to generate one token at a time, making generation sensitive to GPU memory movement and sequential execution. This study evaluates two-token multi-token prediction (MTP) against autoregressive decoding in a controlled single-request deployment on an NVIDIA A10G GPU. A 360-request benchmark covered plain-text, reasoning-intensive, and tool-calling workloads, while runtime telemetry, Nsight Systems, PyTorch Profiler, and selected Nsight Compute measurements were used to explain the observed performance. MTP increased output throughput by \(1.91\times\) to \(2.19\times\) across all prompts and reduced time to first output by 10.0--14.2\%. Median mean acceptance length ranged from 2.370 to 2.595 tokens per verification iteration. Profiling showed that MTP introduced a longer and more complex execution path, including proposal, sampling, attention, gathering, and reduction operations. However, it required 56.4--78.1\% fewer executions of the selected repeating CUDA Graph per generated token. The dominant MTP GEMM kernel was not faster than the dominant autoregressive GEMV kernel, and selected instances of both approached the A10G memory-bandwidth limit. These results show that MTP improved inference through amortization: greater token progress reduced repeated GPU execution sufficiently to outweigh the additional speculative-execution cost.

The development of state-of-the-art large language models is commonly understood as a two-stage process involving pre-training and post-training. We point out the need for an additional intermediate stage called reinforcement mid-training with potential for strong performance gains. In this paper, we formally define the problem and identify three key challenges: (1) inefficient training due to excessive reasoning steps, (2) disregard of the imbalanced token entropy distribution, and (3) underutilization of token information. To address these challenges, we propose RMT, a framework for efficient, adaptive, and unified reinforcement mid-training with various innovative components. In particular, we first introduce a dynamic token budget mechanism that constrains unnecessary reasoning steps and mitigates model overthinking. Next, we design a curriculum-based adaptive sampling method that fosters a progressive learning trajectory from easy to hard tokens. Finally, we present a dual training strategy that combines reinforcement learning with next-token prediction, ensuring targeted learning on key tokens and full exploitation of all token information. Extensive experiments demonstrate the superiority of RMT over state-of-the-art methods, achieving up to +64.91% performance improvement with only 21% of the reasoning length in language modeling. We also show that checkpoints obtained after reinforcement mid-training can benefit the subsequent post-training, yielding up to +18.76% improvement in the mathematical domain.

Later tokens are costlier to generate than earlier tokens. On frontier open weight models, the disparity in cost may rise to as much as 7x. Through frequent compaction, we show across a broad range of long-horizon and long-input tasks that frontier language models sacrifice little to no task accuracy when maintaining short contexts. We further demonstrate that LLM performance does not collapse on long-horizon software engineering tasks even when taking at most one action per context, and based on that we outline the concept of semantic state space models (SSSMs). Our findings question the need for long-context language models.

LLM search agents increasingly rely on tools at inference time, but their trajectories are often constrained by hard limits on both tool calls and generated tokens. Under such dual budgets, better answers require not only stronger models, but also explicit control over which search action should receive the next budget unit and when the accumulated evidence is sufficient to commit a final answer. We study this problem in multi-hop question answering (QA) and formulate it as two-stage inference-time budget control. At search time, our controller assigns each feasible action a task-level Value-of-Information (VOI) score, defined as an operational estimate of marginal task value per unit budget under the current search state and remaining dual budget, and uses this score to choose among retrieval, decomposition, and answer commitment. After search, a selective evidence-grounded finalizer compares the trajectory answer with a refined candidate and rewrites only when the residual error appears to be a low-risk answer-form error. Across four multi-hop QA benchmarks, three LLM backbones, and four budget levels, the method yields positive aggregate gains over four audited baselines under the same hard dual-budget protocol. Ablations show that search-time budget control, especially budget-dependent penalty, provides the main performance gain, while answer-time control helps mainly when the retrieval path is already adequate. These results suggest that inference-time budget control for LLM search agents should govern both how budget is spent during search and how the final answer is committed.

Tree-search decoding is an effective form of test-time scaling for large language models (LLMs), but real-world deployment imposes a fixed per-query token budget that varies across settings. Existing tree-search policies are largely budget-agnostic, treating the budget as a termination condition, which can lead to late-stage over-branching or premature termination. We propose {Budget-Guided MCTS} (BG-MCTS), a tree-search decoding algorithm that aligns its search policy with the remaining token budget: it starts with broad exploration, then prioritizes refinement and answer completion as the budget depletes while reducing late-stage branching from shallow nodes. BG-MCTS consistently outperforms budget-agnostic tree-search baselines across different budgets on MATH500 and AIME24/25 with open-weight LLMs.

Recent work has demonstrated that pre-training in-domain language models can boost performance when adapting to a new domain. However, the costs associated with pre-training raise an important question: given a fixed budget, what steps should an NLP practitioner take to maximize performance? In this paper, we view domain adaptation with a constrained budget as a consumer choice problem, where the goal is to select an optimal combination of data annotation and pre-training. We measure annotation costs of three procedural text datasets, along with the pre-training costs of several in-domain language models. The utility of different combinations of pre-training and data annotation are evaluated under varying budget constraints to assess which combination strategy works best. We find that for small budgets, spending all funds on annotation leads to the best performance; once the budget becomes large enough, however, a combination of data annotation and in-domain pre-training yields better performance. Our experiments suggest task-specific data annotation should be part of an economical strategy when adapting an NLP model to a new domain.

Recent innovations in generative large language models (LLMs) have made their applications and use-cases ubiquitous. This has led to large-scale deployments of these models, using complex, expensive, and power-hungry AI accelerators, most commonly GPUs. These developments make LLM inference efficiency an important challenge. Based on our extensive characterization, we find that there are two main phases during an LLM inference request: a compute-intensive prompt computation, and a memory-intensive token generation, each with distinct latency, throughput, memory, and power characteristics. Despite state-of-the-art batching and scheduling, the token generation phase underutilizes compute resources. Specifically, unlike compute-intensive prompt computation phases, token generation phases do not require the compute capability of the latest GPUs, and can be run with lower power and cost. With Splitwise, we propose splitting the two phases of a LLM inference request on to separate machines. This allows us to use hardware that is well-suited for each phase, and provision resources independently per phase. However, splitting an inference request across machines requires state transfer from the machine running prompt computation over to the machine generating tokens. We implement and optimize this state transfer using the fast back-plane interconnects available in today's GPU clusters. We use the Splitwise technique to design LLM inference clusters using the same or different types of machines for the prompt computation and token generation phases. Our clusters are optimized for three key objectives: throughput, cost, and power. In particular, we show that we can achieve 1.4x higher throughput at 20% lower cost than current designs. Alternatively, we can achieve 2.35x more throughput with the same cost and power budgets.

Developers and consumers increasingly choose reasoning language models (RLMs) based on their listed API prices. However, how accurately do these prices reflect actual inference costs? We conduct the first systematic study of this question, evaluating 8 frontier RLMs across 9 diverse tasks covering competition math, science QA, code generation, and multi-domain reasoning. We uncover the pricing reversal phenomenon: in 21.8% of model-pair comparisons, the model with a lower listed price actually incurs a higher total cost, with reversal magnitude reaching up to 28x. For example, Gemini 3 Flash's listed price is 78% cheaper than GPT-5.2's, yet its actual cost across all tasks is 22% higher. We trace the root cause to vast heterogeneity in thinking token consumption: on the same query, one model may use 900% more thinking tokens than another. In fact, removing thinking token costs reduces ranking reversals by 70% and raises the rank correlation (Kendall's τ ) between price and cost rankings from 0.563 to 0.873. We further show that per-query cost prediction is fundamentally difficult: repeated runs of the same query yield thinking token variation up to 9.7x, establishing an irreducible noise floor for any predictor. Our findings demonstrate that listed API pricing is an unreliable proxy for actual cost, calling for cost-aware model selection and transparent per-request cost monitoring.

Large Language Models (LLMs) achieve impressive reasoning capabilities at the cost of substantial inference overhead, posing substantial deployment challenges. Although distilled Small Language Models (SLMs) significantly enhance efficiency, their performance suffers as they fail to follow LLMs' reasoning paths. Luckily, we reveal that only a small fraction of tokens genuinely diverge reasoning paths between LLMs and SLMs. Most generated tokens are either identical or exhibit neutral differences, such as minor variations in abbreviations or expressions. Leveraging this insight, we introduce **Roads to Rome (R2R)**, a neural token routing method that selectively utilizes LLMs only for these critical, path-divergent tokens, while leaving the majority of token generation to the SLM. We also develop an automatic data generation pipeline that identifies divergent tokens and generates token-level routing labels to train the lightweight router. We apply R2R to combine R1-1.5B and R1-32B models from the DeepSeek family, and evaluate on challenging math, coding, and QA benchmarks. With an average activated parameter size of 5.6B, R2R surpasses the average accuracy of R1-7B by 1.6x, outperforming even the R1-14B model. Compared to R1-32B, it delivers a 2.8x wall-clock speedup with comparable performance, advancing the Pareto frontier of test-time scaling efficiency. Our code is available at https://github.com/thu-nics/R2R.

Commercial large language models bill, scale latency, and budget context per token. Yet tokenizers assign more subword tokens to the same meaning in some languages than in others, so speakers of languages with high token-fertility pay a structural penalty before a model is ever invoked. This penalty is documented for multilingual settings in general, but it has not been measured systematically for African languages at the level of enterprise deployment economics and cognitive context capacity. We measure it across 20 African languages spanning five language families and three scripts (Latin, Ge'ez/Ethiopic, N'Ko; 19 appear in the primary FLORES-200+ corpus, with Nigerian Pidgin measured via MAFAND-MT only), using parallel corpora so that the language effect is isolated from content. Across 11 frontier and open tokenizers on FLORES-200+, every African language carries a tokenization premium above English (median 1.88x on GPT-5 / o200k_base, up to 8.92x for N'Ko); the penalty is largest for Ethiopic and N'Ko scripts (reaching 7-9x) and is near-invariant across corpora (FLORES vs SIB-200 Pearson r = 0.9998). Translated into deployment terms, this results in up to 8.9x inference cost and an equivalent generation-latency multiplier (N'Ko vs English on GPT-5; 7.4x for Amharic), and as little as 11% of English's effective context window. The best currently available tokenizer for African languages, Gemma 4, reduces the mean premium from 3.31x (cl100k_base) to 2.38x, but no tokenizer eliminates the penalty. We release an open measurement tool (afri-fertility), a public leaderboard, a results dataset, and mitigation guidance for African builders. The penalty falls hardest on the languages whose speakers can least afford it, a digital divide encoded directly into the subword vocabulary.

Test-time scaling has become a dominant paradigm for improving LLM agent reliability, yet current approaches treat compute as an abundant resource, allowing agents to exhaust token and tool budgets on redundant steps or dead-end trajectories. Existing budget-aware methods either require expensive fine-tuning or rely on coarse, trajectory-level heuristics that cannot intervene mid-execution. We propose the Budget-Aware Value Tree (BAVT), a training-free inference-time framework that models multi-hop reasoning as a dynamic search tree guided by step-level value estimation within a single LLM backbone. Another key innovation is a budget-conditioned node selection mechanism that uses the remaining resource ratio as a natural scaling exponent over node values, providing a principled, parameter-free transition from broad exploration to greedy exploitation as the budget depletes. To combat the well-known overconfidence of LLM self-evaluation, BAVT employs a residual value predictor that scores relative progress rather than absolute state quality, enabling reliable pruning of uninformative or redundant tool calls. We further provide a theoretical convergence guarantee, proving that BAVT reaches a terminal answer with probability at least 1-ε under an explicit finite budget bound. Extensive evaluations on four multi-hop QA benchmarks across two model families demonstrate that BAVT consistently outperforms parallel sampling baselines. Most notably, BAVT under strict low-budget constraints surpasses baseline performance at 4times the resource allocation, establishing that intelligent budget management fundamentally outperforms brute-force compute scaling.

Multi-teacher on-policy distillation (M-OPD) has emerged as a promising paradigm for consolidating domain-specialized reinforcement learning (RL) experts into a single generalist student via dense, token-level reward supervision. Despite its practical success, the optimization dynamics governing multi-teacher capability integration remain poorly understood, and open, rigorously reproducible recipes are conspicuously lacking. In this work, we establish a controlled M-OPD benchmark on SmolLM3-3B-Base with oracle routing, isolating capability integration from routing ambiguity. Our investigation reveals a pronounced capability integration gap: standard M-OPD captures only 35.6% of the available headroom relative to a domain-routed oracle ensemble, with concise tasks such as instruction following suffering severe degradation and premature stagnation. Crucially, we show that this failure stems not from gradient conflict, but from a severe misallocation of the token-level optimization budget. This pathology is driven by three orthogonal factors: structural sequence-length disparities across domains, dynamic convergence drift due to non-uniform learning rates, and multi-step reward staleness from asynchronous policy updates. To resolve these imbalances, we introduce Open-MOPD, a principled framework incorporating token-share balancing, gap-aware dynamic budget allocation, and student reward refresh. Together, these mechanisms systematically restore cross-domain balance, elevating headroom recovery from 35.6% to 83.4% in a single deployable student. We fully open-source our end-to-end post-training recipe, training trajectories, and evaluation suites on an academically accessible hardware budget.

Deep learning practitioners often operate on a computational and monetary budget. Thus, it is critical to design optimization algorithms that perform well under any budget. The linear learning rate schedule is considered the best budget-aware schedule, as it outperforms most other schedules in the low budget regime. On the other hand, learning rate schedules -- such as the 30-60-90 step schedule -- are known to achieve high performance when the model can be trained for many epochs. Yet, it is often not known a priori whether one's budget will be large or small; thus, the optimal choice of learning rate schedule is made on a case-by-case basis. In this paper, we frame the learning rate schedule selection problem as a combination of i) selecting a profile (i.e., the continuous function that models the learning rate schedule), and ii) choosing a sampling rate (i.e., how frequently the learning rate is updated/sampled from this profile). We propose a novel profile and sampling rate combination called the Reflected Exponential (REX) schedule, which we evaluate across seven different experimental settings with both SGD and Adam optimizers. REX outperforms the linear schedule in the low budget regime, while matching or exceeding the performance of several state-of-the-art learning rate schedules (linear, step, exponential, cosine, step decay on plateau, and OneCycle) in both high and low budget regimes. Furthermore, REX requires no added computation, storage, or hyperparameters.

Language model pretraining has become almost synonymous with prohibitive cost, placing it out of reach for much of the academic and open-source communities. Although strong open-source efforts already exist, including open-weight models and open-source training recipes, a cost-efficient, hardware-accessible, and open-source pretraining recipe has long been missing. Even at a small scale, training Llama-3.2-3B costs over \1.5M, and reproducing SmolLM3-3B needs over 700K. In this report, we present an open pretraining recipe designed to lower this barrier. Using this recipe, we train a collection of Puro-2B models from scratch on up to 1.4 trillion tokens with FP8 precision on consumer-grade RTX 5090 GPUs. The models in the collection differ in token budgets and selected recipe variants. Our best model is trained at a compute cost of less than \6.9K and approaches Qwen2.5-1.5B performance under our evaluation protocol. This cost efficiency is enabled by a combination of approaches, including hardware selection, low-precision training, hyperball optimization, curriculum model averaging, and the data recipe. Beyond the recipe itself, we provide two additional results. First, across the Puro-2B collection, we derive a Puro Cost Scaling Law that relates training cost to average model performance; the fitted law suggests that about 4.4K, less than \$5,090, is sufficient to reach the performance of Qwen2-1.5B. Second, as an end-to-end case study, we examine how pretraining data curricula shape downstream performance after post-training. Such controlled studies are enabled by having access to the full pretraining pipeline rather than model weights alone. We release the full training recipe for Puro-2B, including data, code, and model weights under Apache 2.0 at https://huggingface.co/collections/thu-pacman/puro-2b.

As large language models scale, their training-token budgets must also increase to maintain an appropriate tokens-per-parameter ratio (\(TPP\)). However, high-quality domain data is much harder to scale than general web data. As model size and the training-token budget increase, its fraction in the training mixture tends to decrease. Repeating the available high-quality data provides an effective way to counteract this dilution, but excessive repetition may lead to overfitting. We study this trade-off under practical LLM scaling, where the training-token budget grows proportionally with model size. For a fixed domain, we first find that, surprisingly at a fixed \(TPP\), the optimal repetition count mildly increases with model size. Across different domains, we find that the optimal repetition count is strongly negatively correlated with the final validation loss of a domain: domains with lower loss can generally benefit from more repetitions. In contrast, the amount of unique domain data is only weakly related to the optimal repetition count. These findings suggest that repetition counts tuned on smaller proxy models with the same \(TPP\) can provide a practical estimate for larger models.

A transformer language model performs a bounded amount of computation per token, and recent work by Vishal Sikka, former CEO of Infosys, argues that this bound limits which tasks a model can carry out or verify (arXiv:2507.07505). We ask how much of the budget beneath that ceiling is spent on work the model has already done. Serving is stateless across requests: a model that answers a second question about a document recomputes the document's attention state from the first token. On seven real-world datasets, 98.7% of prompt tokens were text the model had already read. We present Galahad, a memory layer for vLLM, SGLang and llama.cpp that makes this reading a one-time cost. Taliesin saves the model's key-value (KV) state for a block of text and loads it on the next request that contains the same bytes, instead of recomputing it. Blaise keeps the documents themselves and passes the model only the section a question needs. On a recall test with 100 facts hidden in a 97,000-token corpus (Gemma 4 31B), Taliesin alone let the model attend to the whole corpus and answered 98 of 100 on llama.cpp at 3.0 s and 572 J per question, against 10 of 100, 9.3 s and 2,754 J for the same model without Galahad, which could hold only the last 12,000 tokens. With Blaise added, the model read about 668 tokens per question and answered 100 of 100 on all three runtimes at 0.59-0.64 s and 200-213 J; a tuned RAGFlow pipeline answered 77. Storing the corpus is a one-time cost of about 100 s and 28 kJ, whose energy is recovered after 13 questions. Restored state is bit-identical: all 262,144 output logits matched after restart, rehydration and hot-load. Galahad worked with all 30 models we tested under vLLM, and it fails closed: any load that does not pass its checks is recomputed. Together these results move LLM serving from stateless to stateful inference.

Large language models (LLMs) owe much of their stellar performance to expansive input contexts, yet such verbosity inflates monetary costs, carbon footprint, and inference-time latency. Much of this overhead manifests from the redundant low-utility tokens present in typical prompts, as only a fraction of tokens typically carries the majority of the semantic weight. We address this inefficiency by introducing FrugalPrompt, a novel prompt compression framework for LLMs, which retains only the most semantically significant tokens. Leveraging two state-of-the-art token attribution methods, GlobEnc and DecompX, we assign salience scores to every token in an input sequence, rank them to preserve the top-k% tokens in their original order, and obtain a sparse frugalized prompt. We evaluate the approach across four NLP tasks: Sentiment Analysis, Commonsense QA, Summarization, and Mathematical Reasoning, using a suite of frontier LLMs. For the first three tasks, a 20% prompt reduction incurs only a marginal loss in task performance, demonstrating that contemporary LLMs can reconstruct elided context from high-salience cues. In contrast, performance on mathematical reasoning deteriorates sharply, reflecting a stronger dependence on complete token continuity. Further analysis with bottom-k% and random-k% tokens reveals asymmetric performance patterns that may suggest potential task contamination effects, wherein models may resort to shallow memorized patterns from pretraining exposure for conventional NLP tasks. We posit that our work contributes to a more nuanced understanding of LLM behavior in performance-efficiency trade-offs, and delineate the boundary between tasks tolerant to contextual sparsity and those requiring exhaustive context. Our source code and models are available at: https://github.com/Starscream-11813/Frugal-ICL.

A robot's flash endurance is a non-renewable stock: every persisted write spends one of a few thousand program/erase cycles and never refills, yet no fielded robot memory system prices which memories are worth an erase cycle. We treat embodied memory as depreciating capital and price that stock with a single endurance shadow price η, which makes cost-minimizing placement across a RAM / on-board NVM / cloud hierarchy a threshold in a wear-augmented per-byte index. The index is cost-optimal whatever the sign of the value-write association χ; only when χ> 0 does the optimum turn non-monotone, sending a robot's most valuable memories off its flash. The pivot is thus empirical, and we measure χ on real robot logs at a pre-specified gate: its sign is a property of the deployment regime -- positive on recurrent long-horizon manipulation (hatχ approx +1.0 times 10^{-3}, replicated at full power), null on a shorter-horizon suite, and negative on non-recurrent teleoperation. Two boundaries scope the result. The endurance budget is dormant on premium 3,000-P/E TLC at datasheet prices and binding on the commodity QLC/eMMC (sim1,000 P/E) that cheaper edge robots run. And where it binds, a learned wear-aware controller only ties price-based routing on task value, because realized value is tier-invariant across RAM, NVM, and cloud: the rent governs device lifetime and cost, not task performance. Whether wear-aware placement improves task value remains open -- χ is measured against a value proxy, and the non-monotone optimum, while proven, is not yet observed in data.

Enterprises connect language models to their own data through retrieval. The benchmarks that rank multi-hop retrieval systems leave out two facts a buyer needs before a published number can be used: whether the retrieval backbone may be deployed commercially, and what it costs to build. On licensing: the field's dense-retrieval anchor, NV-Embed-v2, is licensed cc-by-nc-4.0. Of the four leading MuSiQue systems we audit (HippoRAG-2, PropRAG, SAG, KET-RAG), three depend on it for their best numbers and none says so. On performance: we measure thirteen embedders from eight makers on one identical MuSiQue harness with bootstrap confidence intervals throughout. Until mid-2026 there was a real commercial tax: the best commercially-licensed embedder trailed the anchor by 2.31 Recall@5 points (95% CI [0.91, 3.71], p=0.001). NVIDIA's Nemotron-3-Embed-8B, released 2026-07-16, has closed it: +0.24 at Recall@5 (95% CI [-0.94, +1.43], p=0.69), -0.58 at Recall@10 (p=0.28). It matches the anchor, does not beat it, and is the only entrant that is commercially licensed, free to self-host, and indistinguishable from the anchor; every other entrant meeting the first two conditions sits 5.2 to 14.6 points below. The durable finding is the paid-versus-free divide: API embedders charge per token on every re-index, self-hosted ones charge nothing. On cost: three of five audited systems (adding Microsoft's GraphRAG) do not disclose indexing cost, and the only published GraphRAG dollar figures span 11x inside one third-party paper (USD 2.30 vs USD 24.94 to index a 5.64 MB corpus once); extrapolated to 1 TB that undisclosed choice separates roughly USD 428K from $4.6M. Our cost model keeps one-time embedding apart from recurring answering: at 1 TB, embedding sits 7.5x-900x below graph construction, and a year of answering at 10,000 queries/day sits 350x or more below it.

LLM agents have become a new class of software user, but every surface they work through was designed for someone else. Pages are built for human eyes, which can skim and ignore; tool schemas for programs, which pay nothing to carry definitions they never call. An agent has neither luxury: it re-reads, and pays again for, everything it is shown on every turn. We present String, an open-source runtime that gives this user an interface of its own and treats the job as an operating-systems problem. Tool knowledge moves out of the agent's context and into a common layer that renders it back one view at a time as Markdown. A single SFMD (String-Flavored Markdown) document declares an application's views, typed actions, navigation, and credentials, and the runtime handles discovery, validation, execution, state, and secrets behind two core verbs: /open to see and /act to do. Web and app turn out to be two renderings of one architecture: an SFMD site serves styled HTML to browsers and the raw document to agents, so one grammar reaches apps, files, shells, and the web, even legacy HTML, with no per-site integration. Views stay partial by design, and the staging is causal: disclosing one tier of detail a single turn too early costs up to 23 accuracy points, while proper staging drops wrong-action selection from 28% to 2%. Privilege follows provenance: a remote page may call HTTP but never the shell, and caller-supplied text never expands a stored secret. On an 87-task benchmark that pairs each task with curated skills, operationalizing those procedures as on-demand String apps yields comparable aggregate success across six models from frontier to small (+1.3pp) while using 33.5% fewer tokens among completed episodes, and the resident interface stays a constant 53 tokens at any catalog size. We report the design, the evaluation, and what three months of production use taught us.

Masked diffusion language models (MDLMs) generate text by iteratively unmasking tokens, but their standard decoder reduces each step to a binary action: a position is either committed to a single token or left fully masked, discarding rich predictive information rather than carrying it forward, and forcing premature, irrevocable commitments that lead to poor performance under a limited decoding budget. In this paper, we reinterpret mask prediction as a clean-state prediction (x-prediction) and show that it can be used to induce a continuous flow in the input embedding space. Building on this view, we propose a continuous decoding framework for MDLMs where tokens can accumulate partial progress at each diffusion step and remain revisable. To match the uneven contextual constraints across positions in language, we replace the globally synchronous schedule in image diffusion with a confidence-based asynchronous update in which the diffusion progress is token-wise accumulated. Additionally, we introduce a lightweight policy network and formulate its training as a reinforcement learning problem. Applied to pretrained LLaDA, our decoder retains 83--97% of full-budget accuracy using under 15% of the diffusion steps, largely outperforming discrete mask-prediction decoding at matched budgets.

Modern agentic AI systems combine multiple large language model agents with heterogeneous skills, yet most architectures either fix communication in advance or allow full broadcast. Both can be inefficient because token cost, latency, redundancy, and error propagation increase with the number of active agents and communication links. We model agent selection and communication as a cooperative game with task-conditioned net utility U(Cmid x)=V(Cmid x)-sum_{iin C}c_i, separating coalition-level costs from agent activation costs. We propose a marginal-value activation rule and greedy router, extend the model to optimize communication edges with per-edge costs, and use estimated Shapley values to predict which agents are worth contacting before and during execution. We connect the problem to submodular maximization and prove two limited guarantees: a curvature-refined bound for a monotone, cardinality-constrained special case, and a tight 1/2-approximation, with a correction for signed objectives, for an unconstrained non-monotone case via double greedy. Neither guarantee applies directly to the main router, which remains a heuristic. We also prove a Shapley-submodularity sandwich bound linking the error of marginal-value routing to a per-agent diminishing-returns quantity. In synthetic experiments, greedy routing achieves 99.5% of brute-force-optimal utility while activating 1.96 of 8 agents on average, compared with 38.8% for full broadcast. Performance is robust to activation cost and redundancy weight but falls to 66% under strong violations of submodularity or noisy value estimates. We distinguish the framework from Shapley pricing, hedonic coalition formation, and communication-graph pruning, and propose evaluation on real multi-agent LLM benchmarks.

A coding agent works on repositories that can be far larger than its context window can hold. Prompt compressors and retrievers offer solutions to this problem, but patching an issue needs two things at once: the files that must change, and their code exactly as written, since the model must reproduce the lines it edits. Retrievers rank code by its similarity to the issue rather than by whether it must change, and compressors delete text to fit the budget. We show that at an equal token budget, an agent resolves more issues when the budget goes to choosing which files to show verbatim, as our simple selector CQ2-CARE does, than when it goes to pruning tokens or retrieving chunks. We fit every context to the same per-issue budget on 300 recent SWE-rebench issues and grade each patch with the benchmark's tests. With GPT-5.6 and Claude Opus 5, CQ2-CARE resolves 43.7% and 46.7% of the issues, 6.3 points more than the best of three compressors with either model, and 9.3 and 12.7 points more than the best of three retrievers. CQ2-CARE even resolves more issues than two methods trained for issue resolution, SweRank (by 0.7 and 1.7 points) and SWE-Pruner (by 2.0 and 7.3). Our analysis shows the defects of the baseline methods: retrievers often miss the files to edit, and compressors keep them but alter their text, so fewer patches apply. Exact-file selection needs no compression model and shows every line as the repository has it, so the model can quote the code it edits.

Every public LLM cost calculator we surveyed treats GPU utilization as a fixed input -- entered by the user, baked in as a preset, or silently assumed at 100% -- never measured against the operator's actual load. We show that this assumption is the dominant source of error: on identical H100 hardware, effective cost spans \0.21 to 15.25 per million output tokens, an underutilization penalty of 2.5-24x across low-to-moderate enterprise loads (1-10 rps) and up to 36.3x near idle -- driven by one operator-controlled variable, offered request rate lambda, which sets in-flight concurrency via Little's Law and which no open-source calculator exposes. Because calculators take utilization as a user-supplied input, any utilization-naive estimate understates true cost by exactly 1/U, systematically mispricing self-hosting -- most severely over-selling it for low-traffic workloads. We propose a measurement methodology that parameterizes the relationship as C_eff = f(H, M, Q, lambda, L), validate it with 42 benchmarks across dense, ultra-sparse MoE, and sparse MoE models, and release vllm-cost-meter, an open-source cost meter that attaches to a live vLLM server and reports real \$/M-tokens against the operator's own traffic. We further show that FP8 quantization benefits the MoE architectures we tested roughly 2.2-2.4x more than the dense model (+69 to +74% vs. +31% peak throughput; n=3, broader validation needed), and our data are consistent with active parameter count, not total model size, being a primary predictor of saturation economics. To rule out single-hardware confounding we repeat the core sweep on A100 80GB PCIe (56 runs): the load-driven spread reproduces at 7.0-11.4x, the active-parameters ordering survives at FP8, and the dense-FP8 advantage inverts on silicon without native FP8 tensor cores -- a hardware-conditional caveat the framework already accommodates.

Tool-using large language model agents frequently replan, retry failed operations, delegate tasks, and resume after crashes. These behaviors can cause one user authorization to be requested and executed multiple times under freshly issued token identifiers, even when each individual token is single-use. We call this failure semantic replay: exceeding the execution budget of a token-independent authorization instance rather than merely reusing an old token identifier. We show that identifier-local token consumption cannot prevent fresh reissuance unless the issuer retains monotonic durable state over the authorized action, confirmation event, and remaining execution budget. We introduce CapLease, an authorization-consumption layer that follows proposal- and authority-level defenses, binds an authenticated user confirmation to a canonical action, and enforces transactional Issue-Prepare-Commit transitions. Across LLM-agent replanning, retry, delegation, concurrency, confirmation-replay, and crash-recovery scenarios, identifier-local tokens permit fresh semantic reissuance, whereas CapLease and an equally stateful Server Ledger prevent duplicate admission and, with an idempotent sink, duplicate external effects. Our results identify durable authorization state, rather than token representation alone, as the systems requirement for replay-resistant agent execution.

Selecting intermediate features from a frozen diffusion model for discriminative tasks is non-trivial because the useful signal is distributed across the denoising trajectory, network depth, and channels. In practice, this requires choosing a timestep, a block, and a channel subset. Existing methods either fix these choices by hand or optimize only a few extraction hyperparameters, most often the timestep, then reuse the same rule across inputs and budgets. We introduce Policy-Mask, a budget-conditioned policy trained on the downstream objective that jointly selects all three under a prescribed channel budget. Without normalization, the channel-score contribution grows with the number of selected channels and dominates the timestep and block updates. We normalize it by subset cardinality, keeping the three policy heads on comparable scales and stabilizing joint training. At matched budgets, Policy-Mask improves over fixed extraction and feature-filter baselines across segmentation, classification, and semantic correspondence, reaching 38.9 mIoU on ADE20K, 75.4% top-1 on CIFAR-100 at 128 dimensions, and 76.3 PCK@0.1 on SPair-71k. Component-wise ablations show that block selection materially affects downstream accuracy, and the highest recorded channel-budget score occurs at 256 retained channels, identifying a practical operating point for compact diffusion representations. We further adapt Policy-Mask to diffusion transformers, including PixArt-Σ and SD3.5 MMDiT, by replacing U-Net channels with image-token hidden dimensions.

APort Vault is a benchmark for payment authorization in tool-using AI agents. It replays 4,371 attacks written by humans against a live payment agent during a public capture-the-flag event, across 14 models from 8 labs, five policy configurations and two replay tracks, with and without a deterministic pre-action check implementing the Open Agent Passport (OAP) specification. 225,964 evaluations completed. We report five distinct events per evaluation, because collapsing them is how an agent benchmark produces a number that does not survive review. Requests are common and their rate differs far more across configurations than across models, though each attack exists at exactly one configuration so policy and attack cohort vary together: 10.9% of model-alone evaluations at Level 1, 3.0% at Level 2, 0.1% at Level 3, 79.4% at Level 4. On the 1,293 Level 4 prompts, each evaluated on every model, request rates run from 71.2% to 84.3%, and 809 prompts (62.6%) elicited a request from all fourteen models, each ending in a successful payment to the level's allowlisted recipient. The authorization boundary is where the conditions diverge. At Levels 2 to 4, transfers to recipients the passport did not permit number 140 of 76,842 with the model alone and 0 of 69,297 behind the layer, and 105 against 0 on 68,970 matched model, prompt and track triples. The zero spans 790 source sessions, giving a per-session upper bound of 0.38%. It was not obtained by refusing payments: 25,370 payments executed behind the layer, while the policy denied 187 of the 25,640 transfer calls it evaluated, 148 of them for a forbidden recipient. We release the 225,964 evaluations, the level passports, the scoring code and the analysis script at huggingface.co/datasets/aporthq/vault-benchmark-v1 .

Outcome RL raises pass@1 but not pass@k at large k (Yue et al., 2025). But every pass@1 is scored under a token budget. Every recipe also trains under one. On physics and math problems at fixed token budgets, we compare the pass@1 of 20 released RL models (1.5B to 32B) with that of the thinking models they started from. We ask whether each pass@1 gain is changed behaviour or new knowledge. We divide each change in pass@1 into three parts. Finishing measures answering before the budget runs out. Reliability measures being right once finished. Reach measures problems RL solves that its base answered wrongly in all eight tries, minus problems the base solves that RL answered wrongly in all eight. We find four things. (1) Where the base often hits the token budget before answering, most of the change, gain or loss, is finishing (82% at 4k and 8k). Given more tokens, the base closes most of the finishing gap itself. (2) Where the base has enough tokens, the gain is reliability: the RL model is graded right more often, even on problems both finish. (3) By our definition, RL rarely solves what its base cannot. Reach adds at most 2.1 percentage points on any of our four physics and math sets. We checked 144 physics problems that only the RL model solved. Of these, 3.5% hold up when resampled or read. In a random sample, nine of ten new RL steps a human confirmed also appear in the base's samples. (4) In a small-scale RL run on an 8B model, the gain on new problems is finishing, with reliability flat. Supplying each problem's physical setup was not shown to improve on plain RL.

Visual token reduction is an effective way to accelerate multimodal large language models (MLLMs), but performance deteriorates rapidly under extremely low token budgets. Existing work has explored both visual-token selection and training-based adaptation to reduced visual inputs. We take a step further by asking how a heavily compressed MLLM should learn from the states induced by its own generations. This setting naturally calls for on-policy self-distillation: a heavily compressed model is supervised on the states induced by its own generations, while its full-token counterpart serves as an information-rich teacher. Based on this insight, we propose LT-OPD, a training framework for extreme visual-token reduction. The student rolls out responses with only a small fraction of visual tokens, and a frozen full-token copy of the same MLLM provides distributional supervision along these student-generated trajectories. To stabilize on-policy learning when visual evidence is severely limited, we further introduce a budget-level curriculum that progressively decreases the token budget during training. Across nine benchmarks on Qwen3.5-4B, LT-OPD raises average retained performance under 5% visual-token retention from 68.6% to 82.3%, outperforming training-free, training-based, and reinforcement-learning baselines at the same budget. The gains transfer consistently to Qwen3.5-9B, GLM-4.6V-9B, and LLaVA-OV-1.5-4B. LT-OPD also reduces KV-cache usage by 85.2% and prefill FLOPs by 85.4% without additional inference overhead, demonstrating that on-policy learning can substantially recover capabilities lost to extreme visual-token reduction.

Vision-language models offer a promising path toward automating radiology report generation, but applying them to full 3D CT volumes poses substantial computational challenges. Modern foundation vision encoders (VEs) can produce tens of thousands of vision tokens per scan, making the visual sequence passed to the large language model (LLM) a primary computational bottleneck. Vision-to-language projectors can compress this sequence to reduce computation, but may discard clinically relevant detail; conversely, effective compression can accommodate higher-resolution inputs while keeping the downstream token count fixed. How this vision-token budget should be allocated across input field of view, spatial resolution, and vision-to-language projection therefore remains an open design question. We systematically evaluate four heterogeneous VEs (CNN- and ViT-based), five token-reducing projectors at up to 64x compression alongside a non-reducing MLP projector baseline, and five instruction-tuned LLMs (1.7B--4B) on two large-scale CT report datasets (CT-RATE and Merlin). At matched LLM token budgets, anatomy-guided region of interest cropping is the most consistent strategy, improving clinical macro F1 in 19 of 20 settings by +3.7 points on average for the 3D ViT Primus encoder and +1.1 for the slice-based 2D ViT Curia encoder. Increasing input resolution further is strongly projector-dependent: the PerceiverResampler, paired with higher-resolution Curia features, yields the strongest configuration in the resolution study on both datasets. Our best configurations achieve state-of-the-art clinical macro F1 on the test sets, reaching 49.5 on CT-RATE and 49.0 on Merlin. Code and models will be published upon publication.

Security-agent evaluations commonly measure peak offensive capability under generous inference budgets, emphasizing vulnerability discovery, exploit development, penetration testing, and CTF completion. Such measurements are useful but incomplete: in operational security, every reasoning step, tool call, telemetry query, and enrichment request consumes budget. We evaluate language-model security agents through this cost-success lens on offensive Cybench challenges and defensive Splunk BOTS v1 investigation challenges. Instead of reporting only best-case success, we compare models at fixed cost levels and decompose performance by inference spend and tool spend. Our results show distinct scalingregimes for red- and blue-team tasks. Offensive CTF performance improves with additional test-time compute, and scaled open-weight models can approach frontier proprietary systems while remaining cost-competitive. Defensive SOC investigation does not scale in the same way: success depends more heavily on disciplined tool use, telemetry navigation, and selective enrichment than on raw reasoning budget alone. We argue that security-agent benchmarks should measure economic efficiency and operational fit alongside task success. Cost-aware, SOC-native evaluations provide a clearer picture of which models are practically useful today and where defensive agents still need to improve. We present an interactive website with our results https://evals.frontier.security.

Large language models are increasingly used to author Infrastructure-as-Code (IaC), where a single insecure default can be deployed directly into production. Prior evaluations report raw vulnerability counts for model-generated IaC, but without a human baseline they cannot determine whether models are actually worse than engineers. We introduce GenIaC-SecBench, a benchmark of 100 deployment scenarios stratified by architectural complexity, evaluated across 12 model configurations from four vendors, producing 1,196 IaC artifacts scanned by three independent policy engines (Checkov, Trivy, KICS). Critically, we also scan 634 human-authored IaC templates with the same toolchain, providing the first size-matched human security baseline. Vulnerability density is strongly inverse to artifact size (Spearman ρ= -0.55, p < 10^{-77}), meaning unmatched comparisons measure size rather than security. When matched on declared-resource count, all model configurations fall within 3.21x--3.87x the human vulnerability density, with the gap widening for simpler tasks (4.9x at one resource, 1.4x at twenty or more). We decompose reasoning into standard generation, prompt-engineered chain-of-thought, and vendor extended-thinking APIs. Vendor extended thinking significantly outperforms prompted chain-of-thought (-12.0%, p = 0.0013), while prompted chain-of-thought is indistinguishable from standard generation (-1.3%, n.s.). Token instrumentation shows extended thinking uses under 1\% of the output budget, explaining its bounded effect. Two negative results also emerge: deployability does not correlate with vulnerability (r = 0.158, p = 0.625), and classical complete-case Friedman testing is infeasible for realistic benchmark designs, motivating the Skillings-Mack statistic. All code, data, and regeneration scripts are released.

Large language models pay a well-documented tax on non-English text: the same content costs several times more tokens, and because attention is quadratic in sequence length, far more compute. We ask how much of this tax is removable. Framing the token layer as source coding -- transformer compute is monotone in sequence length, whose per-atom floor is the Shannon rate H/log_2 V, an object already applied to tokenizers in prior work -- we assemble a token-cost ledger that splits each language's cost, at fixed parallel content, into a removable coding redundancy, a residual coding slack, an intrinsic-content term, and an orthogonal, irreducible grapheme-to-phoneme term that governs the multimodal rather than the text cost. On FLORES-200 across eight languages, a production tokenizer costs up to 8.9times more tokens for Indic scripts than for English; a script-matched code trained on 1,012 sentences removes a median 64% of that excess (bootstrap 95\% CI [0.638, 0.647]), and a script-fair information floor shows the intrinsic content differs by under 6% -- the tax is representational, not informational. A constructed code removes 98% of a controlled source's redundancy, and the token tax implies up to 79times attention cost. We are explicit about scope and failure: this is compute-and-memory accounting, not a model-quality claim; we neither measure nor claim the cross-lingual direction of the orthographic term; and our matched code is a conservative small-data demonstration. We contribute the unifying ledger, the removable-versus-intrinsic attribution, and an open one-command harness.

Model routing can reduce the cost of LLM applications by assigning simpler requests to cheaper models without sacrificing the quality of the response. However, most routers treat requests as independent. Stateful agents do not operate under this assumption: a user turn may require multiple model and tool calls, its difficulty depends on the state produced by earlier turns, and changing models can sacrifice reuse of the conversational prompt cache. We study zero-shot, cost-aware routing at the user-turn boundary, systematically varying the routing rule, access to session history, candidate models that differ in size, reasoning effort, or model provider, and prompt-cache behavior. We evaluate on multi-turn versions of SheetCopilot and SpreadsheetBench 2, where persistent workbooks and cross-turn dependencies make isolated routing decisions especially brittle. At a matched offloading rate, giving the router session history improves task pass rate over the same router without history, with gains for every model pair; history-aware routing remains within one point of always using the expensive model. We also found that cost depends on execution dynamics rather than token prices alone: the cheaper model's per-task cost advantage narrows from 6.8x to 3.2x on longer sessions. A closed form cache analysis further shows that downward switches (expensive to cheap) recover their cache cost within two LLM calls, whereas upward switches do not break even under the evaluated conditions. Together, these results show that agent routing is a sequential decision problem: effective policies must account for what happened earlier in the session and for the asymmetric cost of switching models.