new

Get trending papers in your email inbox!

Subscribe

Daily Papers

byAK and the research community

Nov 19

HAMburger: Accelerating LLM Inference via Token Smashing

The growing demand for efficient Large Language Model (LLM) inference requires a holistic optimization on algorithms, systems, and hardware. However, very few works have fundamentally changed the generation pattern: each token needs one forward pass and one KV cache. This can be sub-optimal because we found that LLMs are extremely capable of self-identifying the exact dose of information that a single KV cache can store, and many tokens can be generated confidently without global context. Based on this insight, we introduce HAMburger, a Hierarchically Auto-regressive Model that redefines resource allocation in LLMs by moving beyond uniform computation and storage per token during inference. Stacking a compositional embedder and a micro-step decoder in between a base LLM, HAMburger smashes multiple tokens into a single KV and generates several tokens per step. Additionally, HAMburger functions as a speculative decoding framework where it can blindly trust self-drafted tokens. As a result, HAMburger shifts the growth of KV cache and forward FLOPs from linear to sub-linear with respect to output length, and adjusts its inference speed based on query perplexity and output structure. Extensive evaluations show that HAMburger reduces the KV cache computation by up to 2times and achieves up to 2times TPS, while maintaining quality in both short- and long-context tasks. Our method explores an extremely challenging inference regime that requires both computation- and memory-efficiency with a hardware-agnostic design.

  • 2 authors
·
May 26, 2025

Hierarchical Diffusion Transformer for End-to-End Image Generation

Pixel-space diffusion learns visual representations and image generation jointly in a single end-to-end training stage, without a separately pretrained visual encoder. However, plain pixel-space Transformers still trail leading two-stage latent diffusion models in generation quality and training efficiency. We hypothesize that part of this gap arises from retaining one token per patch throughout the backbone, which does not explicitly encourage a visual hierarchy that summarizes related content at progressively coarser levels. To address this, we introduce Hierarchical Diffusion Transformers (HiT), which progressively group related tokens and represent each group with a single token. The main Transformer bottleneck processes these compact group tokens, and ungrouping restores features to their original image positions. A compression budget regulates the average compression ratio across the batch, allowing group memberships and counts to vary across images and diffusion timesteps during both training and generation. The model exhibits emergent semantic structure, learning to group coherent objects and object parts without segmentation labels. In text-to-image generation, HiT further learns to group caption words with image patches depicting the same concepts, without region annotations. HiT achieves 1.48 FID on ImageNet 256times256 and reaches comparable generation quality to JiT using approximately 40\% of its cumulative training FLOPs. For text-to-image generation, HiT exceeds UniLiP's average score across five benchmarks with roughly 6times fewer denoiser parameters and 3times fewer forward FLOPs. These results suggest that learning a visual hierarchy within the generator breaks down the barrier to stronger image generation without external visual pretraining.

  • 7 authors
·
Nov 16

Distill to Delete: Unlearning in Graph Networks with Knowledge Distillation

Graph unlearning has emerged as a pivotal method to delete information from a pre-trained graph neural network (GNN). One may delete nodes, a class of nodes, edges, or a class of edges. An unlearning method enables the GNN model to comply with data protection regulations (i.e., the right to be forgotten), adapt to evolving data distributions, and reduce the GPU-hours carbon footprint by avoiding repetitive retraining. Existing partitioning and aggregation-based methods have limitations due to their poor handling of local graph dependencies and additional overhead costs. More recently, GNNDelete offered a model-agnostic approach that alleviates some of these issues. Our work takes a novel approach to address these challenges in graph unlearning through knowledge distillation, as it distills to delete in GNN (D2DGN). It is a model-agnostic distillation framework where the complete graph knowledge is divided and marked for retention and deletion. It performs distillation with response-based soft targets and feature-based node embedding while minimizing KL divergence. The unlearned model effectively removes the influence of deleted graph elements while preserving knowledge about the retained graph elements. D2DGN surpasses the performance of existing methods when evaluated on various real-world graph datasets by up to 43.1% (AUC) in edge and node unlearning tasks. Other notable advantages include better efficiency, better performance in removing target elements, preservation of performance for the retained elements, and zero overhead costs. Notably, our D2DGN surpasses the state-of-the-art GNNDelete in AUC by 2.4%, improves membership inference ratio by +1.3, requires 10.2times10^6 fewer FLOPs per forward pass and up to 3.2times faster.

  • 3 authors
·
Sep 28, 2023

You Only Judge Once: Multi-response Reward Modeling in a Single Forward Pass

We present a discriminative multimodal reward model that scores all candidate responses in a single forward pass. Conventional discriminative reward models evaluate each response independently, requiring multiple forward passes, one for each potential response. Our approach concatenates multiple responses with separator tokens and applies cross-entropy over their scalar scores, enabling direct comparative reasoning and efficient N-way preference learning. The multi-response design also yields up to Ntimes wall-clock speedup and FLOPs reduction over conventional single-response scoring. To enable N-way reward evaluation beyond existing pairwise benchmarks, we construct two new benchmarks: (1) MR^2Bench-Image contains human-annotated rankings over responses from 8 diverse models; (2) MR^2Bench-Video is a large-scale video-based reward benchmark derived from 94K crowdsourced pairwise human judgments over video question-answering spanning 19 models, denoised via preference graph ensemble. Both benchmarks provide 4-response evaluation variants sampled from the full rankings. Built on a 4B vision-language backbone with LoRA fine-tuning and a lightweight MLP value head, our model achieves state-of-the-art results on six multimodal reward benchmarks, including MR^2Bench-Image, MR^2Bench-Video, and four other existing benchmarks. Our model outperforms existing larger generative and discriminative reward models. We further demonstrate that our reward model, when used in reinforcement learning with GRPO, produces improved policy models that maintain performance across standard multimodal benchmarks while substantially improving open-ended generation quality, outperforming a single-response discriminative reward model (RM) baseline by a large margin in both training stability and open-ended generation quality.

A Controlled Study of Attention-Only Transformers

Feed-forward networks hold two thirds of a transformer's non-embedding parameters, yet the architecture has not received a necessity test that controls parameters, compute, and depth at once. We pretrain attention-only decoder transformers (Simple Attention Networks, SANs) against standard transformers matched separately for parameter count, training FLOPs, and depth (2 to 48 layers), for up to 105B tokens at 6M to 87M parameters. Deleting feed-forward layers in place is costly: the standard transformer leads by 0.47 nats at matched depth and 0.26 nats at matched FLOPs. Reallocating the freed budget into attention depth closes the gap: at matched parameters the difference is 0.006 nats (0.27 percent of loss), reproducible to one part in ten thousand across seed pairs, shrinking across 5B, 30B, and 105B budgets, and holding near 0.02 nats across a 29x size range. Three measurements localize the remaining gap to parametric recall: attention-only models are better on context-grounded answers and worse where knowledge must come from weights. Weight spectra show why: routing matrices (Q/K) crystallize early, content matrices accumulate rank slowly, and removing feed-forward layers relocates this accumulation to the attention output projection. QK-normalization, not feed-forward layers or residual gating, keeps 48-layer attention-only stacks trainable. The deficit concentrates on low-context query prediction and localizes there entirely by the largest budget. A pre-registered test confirms the account: it predicts a 0.02 to 0.05 nat gap on knowledge-dense web text; a matched pair trained on fineweb-edu measures 0.040. Within the tested regime, attention does the rest.

  • 8 authors
·
Jul 19

TransKD: Transformer Knowledge Distillation for Efficient Semantic Segmentation

Large pre-trained transformers are on top of contemporary semantic segmentation benchmarks, but come with high computational cost and a lengthy training. To lift this constraint, we look at efficient semantic segmentation from a perspective of comprehensive knowledge distillation and consider to bridge the gap between multi-source knowledge extractions and transformer-specific patch embeddings. We put forward the Transformer-based Knowledge Distillation (TransKD) framework which learns compact student transformers by distilling both feature maps and patch embeddings of large teacher transformers, bypassing the long pre-training process and reducing the FLOPs by >85.0%. Specifically, we propose two fundamental and two optimization modules: (1) Cross Selective Fusion (CSF) enables knowledge transfer between cross-stage features via channel attention and feature map distillation within hierarchical transformers; (2) Patch Embedding Alignment (PEA) performs dimensional transformation within the patchifying process to facilitate the patch embedding distillation; (3) Global-Local Context Mixer (GL-Mixer) extracts both global and local information of a representative embedding; (4) Embedding Assistant (EA) acts as an embedding method to seamlessly bridge teacher and student models with the teacher's number of channels. Experiments on Cityscapes, ACDC, and NYUv2 datasets show that TransKD outperforms state-of-the-art distillation frameworks and rivals the time-consuming pre-training method. Code is available at https://github.com/RuipingL/TransKD.

  • 7 authors
·
Feb 27, 2022

Multi-View MoE: Letting Every Router Shape Expert Computation

Sparse mixture-of-experts (MoE) models expand parameter capacity by activating a small subset of experts per token. Expert selection directly affects model capabilities under a fixed active-expert budget. Dense evaluation of all experts reveals weak agreement between single-router rankings and expert-utility evaluation rankings, especially in middle and deep layers. Richer signals for selection and computation incur additional cost: selection based on internal expert activations requires preliminary computation, while processing multiple full representations separately through the same expert repeats its computation. We propose Multi-View MoE to improve expert selection with multiple routers and reuse their preferences to modulate expert computation under a fixed expert budget. Its Multi-Router submodule aggregates sigmoid scores from independently initialized routers. These routers develop highly differentiated expert preferences during training, allowing experts strongly favored by one router or jointly supported by several routers to be selected. Multi-View further uses each selected expert's relative scores across routers to mix learned latent views into its input, reusing routing preferences for both selection and computation. Each selected expert executes one forward pass, preserving the active-expert count, expert FLOPs, and all-to-all communication volume. On a 24-layer 10B-A1B model trained on 708B tokens, full Multi-View MoE improves the nine-task downstream mean by 1.18 percentage points over the baseline. Multi-Router alone yields a 0.6-percentage-point gain and improves the rank correlation between router scores and evaluated expert utility, with relative increases of 8.1% globally and approximately 58% in the layer with the lowest single-router alignment. Controlled mixing interventions show the model uses input differences that depend on both the expert and the token.

  • 5 authors
·
Oct 28

Induction Signatures Are Not Enough: A Matched-Compute Study of Load-Bearing Structure in In-Context Learning

Mechanism-targeted synthetic data is increasingly proposed as a way to steer pretraining toward desirable capabilities, but it remains unclear how such interventions should be evaluated. We study this question for in-context learning (ICL) under matched compute (iso-FLOPs) using Bi-Induct, a lightweight data rewrite that interleaves short directional copy snippets into a natural pretraining stream: forward-copy (induction), backward-copy (anti-induction, as a directional control), or a balanced mix. Across 0.13B-1B decoder-only models, we evaluate (i) few-shot performance on standard LM benchmarks and function-style ICL probes, (ii) head-level copy telemetry, and (iii) held-out perplexity as a guardrail. Bi-Induct reliably increases induction-head activity, but this does not translate into consistent improvements in few-shot generalization: on standard LM benchmarks, Bi-Induct is largely performance-neutral relative to natural-only training, while on function-style probes the 1B natural-only model performs best. Despite explicit backward-copy cues, anti-induction scores remain near zero across scales, revealing a strong forward/backward asymmetry. Targeted ablations show a sharper distinction: removing the top 2% induction heads per layer harms ICL more than matched random ablations, with the largest relative drop occurring in the natural-only models. This indicates that natural-only training produces more centralized, load-bearing induction circuitry, whereas Bi-Induct tends to create more distributed and redundant induction activity. Our main conclusion is that eliciting a mechanism is not the same as making it load-bearing. For data-centric foundation model design, this suggests that synthetic data interventions should be evaluated not only by signature amplification, but by whether they create causally necessary computation while preserving natural-data modeling quality.

  • 2 authors
·
Mar 13

AsyncSpade: Efficient Test-Time Scaling with Asynchronous Sparse Decoding

Test-time scaling (TTS) boosts LLM reasoning via long chain-of-thought (CoT), but the linear KV-cache growth amplifies the memory-bound bottleneck of LLM decoding. Query-aware page-level sparse decoding can achieve state-of-the-art performance under constrained FLOPs budgets, but is limited by both sequential-dependent page filtering and coarse-grained token selection, hampering serving efficiency and model performance on TTS tasks under high concurrency and long CoT scenarios (consuming even higher runtime than the forward pipeline itself). In this paper, we first find that the current-step query state can be accurately approximated in a unified manner from a short window of recent queries, enabling training-free query-aware sparsity without waiting in the decoding loop. We propose AsyncSpade, an asynchronous framework for efficient TTS built on two core components: (1) a novel light-weight temporal-regressive module that predicts the next-token query state; (2) an asynchronous and disaggregated framework that decouples the KV cache filtering from the auto-regressive decoding loop, overlapping the token-level KV selection with the forward inference computation through asynchronism. To our knowledge, AsyncSpade is the first to eliminate the sequential dependence without sacrificing model performance. We validate the effectiveness of AsyncSpade on common LLM serving setups with an A100 node, where AsyncSpade fully overlaps KV-cache operations with the inference pipeline, achieving theoretical optimal time-per-output-token (TPOT). Specifically, AsyncSpade delivers over 20% reduction on TPOT compared to SoTA baseline (i.e. Quest) and at least 50% TPOT reduction compared to full attention on Qwen3-8B and Qwen3-32B models, while matching or surpassing their accuracy on various TTS benchmarks (AIME-24/25, GPQA-Diamond, MATH-500).

  • 5 authors
·
Oct 7, 2025

DLLM-JEPA: Joint Embedding Predictive Architectures for Masked Diffusion Language Models

Joint Embedding Predictive Architectures (JEPAs) have reshaped self-supervised representation learning in vision. The recent LLM-JEPA ported JEPA to autoregressive language models but inherited two steep costs from the causal-attention substrate: it demands explicit multi-view data (e.g., text-code pairs), and it requires two gradient-carrying forward passes per step. We introduce DLLM-JEPA, which pairs JEPA with masked-diffusion language models to eliminate both costs at once. The bidirectional attention of diffusion models yields two semantically distinct views of the same input via different masking rates -- no explicit pairs needed -- and supports a single gradient-carrying forward pass, cutting training FLOPs by 33% relative to LLM-JEPA. DLLM-JEPA improves over diffusion-only fine-tuning in every (task, architecture) combination we evaluate: up to +18.7 pp on LLaDA-8B GSM8K and +11.4 pp on Dream-7B GSM8K, with consistent positive gains on Spider, NL-RX-SYNTH, and Django. Beyond accuracy, DLLM-JEPA exhibits a dual-win property: on LLaDA-8B with the Wide-t configuration, it simultaneously raises GSM8K accuracy (67.1 vs. 65.2, +1.8 pp), drives held-out Wikitext loss below the pre-trained base, and preserves MMLU accuracy at base level across three fine-tuning seeds -- whereas an L2-to-base parameter anchor matches baseline accuracy with no task gain. Layer-wise probing reveals the mechanism: a geometric-functional drift dissociation in which the fine-tuned backbone moves further from the pre-trained weights than the baseline yet forgets less on held-out Wikitext, with the amplification concentrated in middle transformer layers. The pattern appears on Dream-7B as well, indicating the phenomenon is not specific to a single backbone.

  • 1 authors
·
May 23

JoVox: Joint Hierarchical Flow Matching for Lightweight Voxel Generation

Generating high-fidelity voxels while keeping their lightweight nature remains an open problem in computer vision. Current generators are either high-fidelity but computationally heavy, or lightweight but of lower fidelity. Hierarchical methods keep the lightweight nature of voxels. However, their coarse levels, which decide where all finer cells can appear, are generated with the least information about the shape and are never revisited, so they are often plausible but wrong. Our key insight is to let finer cells take part in deciding coarse cells. To this end, we propose JoVox, a discrete flow matching framework that learns the joint distribution of all levels of the voxel hierarchy, instead of a separate distribution for each level, so that coarse cells can be predicted with finer cells already generated nearby. Within this framework, we further introduce a two-level prediction, which outputs the cells two levels finer at once instead of one level finer, halving the number of levels that the network processes and yielding voxels of higher fidelity. We also introduce an unmasking strategy that revisits early decisions once finer cells are generated and corrects inconsistencies in masked diffusion models. Our framework outperforms the baselines on both unconditional and image-conditioned voxel generation. Compared with the current state-of-the-art TRELLIS.2, we achieve 8x higher training throughput, 1/15 of the FLOPs in a forward pass, train on 4 vs. 32 GPUs, and obtain higher quality with an 11% lower Chamfer distance.

Cumulative-Goodness Free-Riding in Forward-Forward Networks: Real, Repairable, but Not Accuracy-Dominant

Forward-Forward (FF) training allows each layer to learn from a local goodness criterion. In cumulative-goodness variants, however, later layers can inherit a task that earlier layers have already partially separated. We formalize this phenomenon as layer free-riding: under the softplus FF criterion, the class-discrimination gradient reaching block d decays exponentially with the positive margin accumulated by preceding blocks. We then study three local remedies -- per-block, hardness-gated, and depth-scaled -- that recover current-layer separation measures without relying on backpropagated gradients. On CIFAR-10 and CIFAR-100, these remedies dramatically improve layer-separation statistics, with 4times--45times gains in deeper layers, while changing accuracy by less than one percentage point for non-degenerate training procedures. Tiny ImageNet provides a tougher cross-dataset check for our selected block-wise configuration and reveals the same qualitative gap between layer-health diagnostics and final accuracy. Calibration experiments further show that architecture and augmentation choices have a larger effect on final accuracy than the training-rule modifications studied here. Cumulative free-riding is therefore a real and repairable optimization pathology. Nonetheless, for the FF training rules, architectures, and datasets we study, it is not the dominant factor limiting achievable accuracy.

  • 1 authors
·
May 6

Can We Really Learn One Representation to Optimize All Rewards?

As machine learning has moved towards leveraging large models as priors for downstream tasks, the community has debated the right form of prior for solving reinforcement learning (RL) problems. If one were to try to prefetch as much computation as possible, they would attempt to learn a prior over the policies for some yet-to-be-determined reward function. Recent work (forward-backward (FB) representation learning) has tried this, arguing that an unsupervised representation learning procedure can enable optimal control over arbitrary rewards without further fine-tuning. However, FB's training objective and learning behavior remain mysterious. In this paper, we demystify FB by clarifying when such representations can exist, what its objective optimizes, and how it converges in practice. We draw connections with rank matching, fitted Q-evaluation, and contraction mapping. Our analysis suggests a simplified unsupervised pre-training method for RL that, instead of enabling optimal control, performs one step of policy improvement. We call our proposed method one-step forward-backward representation learning (one-step FB). Experiments in didactic settings, as well as in 10 state-based and image-based continuous control domains, demonstrate that one-step FB converges to errors 10^5 smaller and improves zero-shot performance by +24% on average. Our project website is available at https://chongyi-zheng.github.io/onestep-fb.

  • 3 authors
·
Feb 10

Tree-Ising: Repairing Language Models with Bit-Level Ising Edits

Weight-space programming, introduced in recent MIT work by Takehi et al., enables a frontier LLM agent to read a network's weights and write a repair. Their GPT-6 Astra agent corrected the reasoning loops of Qwen3.5-0.8B without gradients, using 158 agent turns, 89.7 million tokens, and 50 forward-pass requests. Our method, Tree-Ising, casts this repair as a sparse discrete optimization problem over the model's own BF16 weights. Forward statistics define an objective that weakens a loop direction, strengthens a concluding direction, and preserves the layer's output on productive tokens. Within a state-tree representation, where lossless BF16 storage acts as a special case, the admissible edits are sign flips and exponent shifts that preserve all shared group exponents and mantissas. We decompose this objective exactly into independent QUBO (Ising) problems, one per output row, which we solve on a GPU using simulated annealing or quantum-inspired simulated bifurcation. On the agent's two layers, Tree-Ising achieves 54.6% on MATH-500, outperforming the agent's edit (51.7%) under the same evaluation protocol, while the base model scores 7.9% as reported. The simulated bifurcation solver yields a statistically indistinguishable repair at 55.1%. Both solvers outperform the agent edit across all five math benchmarks, with statistically significant improvements on MATH-500 and GSM8K after correcting for multiple comparisons. Our approach modifies 13% to 17% fewer weights and reduces BF16 bit flips by a factor of 2.2 to 2.3. It requires only four forward-pass requests and zero LLM tokens, with the annealing recipe costing about $0.90 in GPU time. On Qwen3.5-2B, with layers and dose selected using only the base model, the same recipe increases greedy GSM8K accuracy from 21.6% to 69.4%. As a proof of concept, we execute an eight-variable repair subproblem using QAOA on an IBM Heron processor.

  • 2 authors
·
Nov 9 7