Daily Papers

Conversational memory changes during use, so endpoint question answering alone cannot establish how a persistent state accumulates, ages, or incorporates revisions. We introduce LSREP, a Longitudinal State-Replay Evaluation Protocol combining ordered replay, explicit lifecycle schedules, repeated probes, evolving reference answers, and mechanism-fidelity checks. Its architectural case study is ICE v2, a local-first memory middleware with typed stores, retrieval fusion, and dynamic context budgets. The private, single-user instantiation contains 1,985 turns, 219 distinct probes, and 1,211 probe-checkpoint observations across 52 checkpoints. On three ordinary-density datasets, ICE v2 has a near-zero mean quality difference from vector-RAG while selecting 32% fewer fragments but using 6.6% more estimated prompt tokens. A fourth, dense dataset exposes catastrophic failures of the unbudgeted baseline. The fidelity audit limits attribution: procedural retrieval is defective, several mechanisms are unexercised, and graph utility is not established. In a complementary matched public diagnostic, ICE v2 loses decisively to pure vector-RAG on LongMemEval: 50.8% versus 72.8% in the evidence-only oracle and 43.0% versus 69.5% in full-S. Paired differences are -22.0 points (95% CI [-26.6, -17.4]) and -26.5 ([-31.3, -21.8]). Conservative abstention accompanies severe multi-session and temporal failures. ICE uses less context in this diagnostic, establishing a quality-cost trade-off rather than superior efficiency. Together, replay, fidelity auditing, and public endpoint testing expose distinct failure modes that neither architectural descriptions nor aggregate scores identify alone.

We introduce Humans-Junior, a 3.8B model that matches GPT-4o on the FACTS Grounding public subset within a pm 5 pp equivalence margin. Results. On Q1--Q500 under identical judges, GPT-4o scores 73.5% (95% CI 69.5--77.2) and Humans-Junior 72.7% (95% CI 68.7--76.5); the paired difference is 0.8 pp (bootstrap 95% CI -3.1 to +4.7; permutation p = 0.72; Cohen's d = 0.023). TOST establishes equivalence at pm 5 pp (not at pm 3 pp). When purchased as managed APIs, Humans-Junior's base model (Phi-3.5-mini-instruct) is approx 19times less expensive than GPT-4o on Microsoft AI Foundry pricing; self-hosted or edge deployments can drive incremental inference cost toward zero. Measured vs estimated pricing sources are tabulated in Appendix E. Method. Our approach combines minimal directed "Exoskeleton Reasoning" scaffolds with behavioral fine-tuning that teaches protocol compliance (epistemic discipline) rather than domain answers. Fine-tuning alone adds little; combined, they synergize (+17.7 pp, p < 0.001) and reduce variance (approx 25%). In prompt-only settings on frontier models (Q1--Q100; non-comparable), directed reasoning improved GPT-4o by +11.8 pp to 85.3% and Gemini-2.5-Pro by +5.0 pp to 93.3% (baseline 88.3%, n = 100); see Section~5. TL;DR. A 3.8B model achieves GPT-4o-level FACTS accuracy (equivalent within pm 5 pp on Q1--Q500). Cloud pricing shows approx 19times lower cost versus GPT-4o, and self-hosted/edge deployments can approach zero marginal cost. Pricing sources are listed in Appendix E. Frontier prompt-only gains (Q1--Q100; non-comparable) and optimized-prompt exploratory results under earlier judges are summarized in Appendix F. Keywords: Small Language Models, Factual Grounding, Directed Reasoning, Fine-Tuning, Model Alignment, Cost-Efficient AI

Retrieval-augmented generation asks one question at volume: does this evidence support this claim? The answer is one bit, yet most pipelines obtain it from a text generator and pay generation prices for the prose wrapped around it. We compare a non-generative decision model, which returns a typed value from a declared enumeration together with a native probability and no text, against a frontier LLM configured as an engineer would deploy it for binary classification. Both answer the same 495 human-labelled claims from LLM-AggreFact, so every comparison is within-item. We report three things. First, the two systems are statistically indistinguishable on the metric this task is conventionally scored by, a paired difference of -0.56 points, Jev minus Astra, at a 187-fold cost difference. Second, that tie carries almost no information about deployment: the systems disagree on 15.2% of claims, and the cheaper one admits 23.2% of the claims annotators marked unsupported against the frontier model's 14.1%. A component whose job is to catch unsupported claims should be judged on that rate, and the conventional headline hides it. Third, because the errors differ they compose: routing the claims the cheap model is confident about and escalating the rest reaches 74.10 at USD 2.92 per thousand claims against 73.83 at USD 8.93, with the threshold chosen out of fold. But a matched random-escalation control at in-sample thresholds, spending the same budget by coin flip, does as well, so what buys the saving is the escalation share and not the probability that picks what to escalate. Measured against the cheap model it is built from, the cascade costs 61-fold more for 0.83 points, 95% CI [-3.36, 4.86], and it is the more permissive guard of the two. Equal accuracy conceals unequal errors. All predictions, prompts, pricing and code are released.

Arabic large language models must refuse harmful prompts without over-refusing benign or sensitive prompts, yet a single refusal rate hides this trade-off. We evaluate it using benign refusal B and harmful-prompt refusal H, where H measures refusal rather than harmful compliance. Across five Arabic-capable models and 130 runs on the full human-written AraSafe set, refusal-only supervised fine-tuning (SFT) collapses toward blanket refusal, whereas selected mixed-SFT configurations reach H = 90% to 93% at B = 14% to 23%; four selected configurations exceed the H = 90% target in all three runs, while Fanar does so in two of three. Direct Preference Optimization (DPO) and inference guards change B and H differently across models rather than acting as uniform upgrades. In a blinded 300-response audit, annotator binary-refusal agreement is 89.0% (kappa = 0.78); Qwen3Guard and Aya Expanse 32B reach 88.7% and 91.0% accuracy, respectively, with no conclusive paired difference. Selected SFT raises H on Arabizi for all five models, but none reaches 90%, showing only partial transfer from Modern Standard Arabic. Overall, the results support model-specific operating-point selection: set a deployment target and retain only interventions that improve it.

Model-based reinforcement-learning agents of the DreamerV3 family forget catastrophically when trained on task sequences, even when an unbounded replay buffer preserves every earlier experience. We ask a question the continual-RL literature has assumed an answer to but never measured: which component forgets? Under never-clear replay, pre-registered component-level probes (n=3 seeds throughout) show that the world model retains essentially everything measurable about old tasks -- reward discrimination (retention ratio ~1.0), value estimates, and termination structure -- while the actor's behavior collapses. Forgetting in this regime is a channel problem, not a memory problem. We demonstrate this by intervention: with the world model frozen and identical imagined rollouts, reinforcement learning in imagination fails to recover a lost skill (0/3 seeds), while supervised self-imitation on the world model's own graded dreams recovers it on 3/3 seeds with zero environment interaction. Interleaved during training, this graded dream rehearsal yields a task-label-free, parameter-constant continual learner: 3/3 four-task chains retained where plain replay passes 0/3, 3/3 eight-task chains, and consistent gains over matched real-episode cloning (paired difference +0.13, bootstrap 95% CI [0.07, 0.24], complete seed separation). The dream-grading step is load-bearing: we characterize two scoring failure modes, provide an offline selection gauge that caught both before they contaminated results, and give a realized-first grading rule that closes them. All experiments were pre-registered with committed protocols; every refuted hypothesis is reported.

Trader-facing dynamic fees are increasingly proposed for automated market makers (AMMs), but historical data do not identify how order flow would respond: trader-facing fees do not vary, trader types are latent, and a replayed tape is not a sequential decision environment. We therefore construct a minimal closed-loop simulator in which the missing signal exists by construction: two constant-product pools repriced by an equilibrium-inspired dynamic-fee rule, fee-sensitive noise flow, and closed-form CEX--AMM arbitrage. Equilibrium is used as a closure principle, not as an object the trader learns. Against a tuned benchmark ladder of schedule, planning, lookahead, and tabular policies, a small DQN is the only evaluated valid policy whose paired improvement over tuned one-step routing excludes zero. On a reserved final block of 1{,}000 seeds with completion forced to 1.0 for every policy, it reduces implementation shortfall under every tested intra-step ordering, by 13.3bps of order notional under the pre-specified agent-last ordering, and the edge is concentrated in, and learned from, dynamic-fee environments: under constant fees the paired difference is indistinguishable from zero. The result is model-conditioned counterfactual evidence about execution control in AMMs, not evidence about historical traders, equilibrium play, or deployable profit.

Modern deep neural networks treat parameters as static floating-point matrices stored in physical memory, incurring Von Neumann memory bottlenecks and representation collapse. We formulate Orbital Error Dynamics (OED), an analytical framework wherein synaptic weights are not stored masses (O(W)), but transient topological resonances (O(1)) derived procedurally from the complex quadratic polynomial map z_{n+1} = z_n^2 + c. We introduce the Bent Sine Wave Hypothesis, demonstrating that non-equilibrium living systems emerge when harmonic waves curl inward through environmental drag toward the cardioid cusp (c = 1/4). We define the Observer Horizon Geometry in parameter space, identifying interior resonance shoulder loci X_upper = (0.25, +0.18) and X_lower = (0.25, -0.18) between the fixed-point basin and the true boundary at c = 0.25 +/- 0.50i. To escape non-convex stagnation without loss zeroing, we introduce a heavy-tailed Biomimetic Perturbed Jump Operator (Omega_tunneling) inspired by mammalian fertilization zinc sparks. We further couple an enteric-cranial Dual-Brain architecture shielded by adaptive CD4+ regulatory immune gating (M_CD4), and project the 4-nucleotide genetic basis (A, T, C, G) across quadrants in C. Multi-seed empirical validation on the Two-Moons manifold (5 seeds, 80/20 train/test split, 32x32 grid, zero test-time updates, zero label leakage) demonstrates that procedural parameterization from a 24-byte coordinate seed achieves 77.67% +/- 5.35% clean test accuracy (within an 8.00-point paired difference of an unconstrained gradient baseline at 85.67% +/- 5.35%, 95% CI: [-1.07%, 17.07%]) and 71.33% +/- 3.80% under distribution shift (N(1.2, 0.4)), alongside conceptual equivalence with an analog optical co-processor.

People look up a song by humming it, but also by whistling it or singing it with the wrong words. Published query-by-humming systems and benchmarks focus on hums and sung queries, and no public dataset pairs whistles or wrong-lyrics singing with song recordings. We study both query types with a melody-only retrieval system: a key-normalized pitch track embedded by a 5.2M-parameter contour encoder, with a spectral-peak tracker for whistles. First, a physics-grounded generator renders synthetic whistles and nonsense-syllable singing from the vocal melodies of 11,083 songs. Training on them raises top-1 retrieval of held-out synthetic queries against an 11,050-song gallery from 0.180 to 0.251 for whistles and from 0.193 to 0.264 for wrong-lyrics singing, but the gain does not transfer: on the MLEnd test performers the best synthetic-only model gets 210 of 339 real whistles right at top-1, against 214 for the baseline. Second, a closed-set fine-tune on real MLEnd whistles with performer-disjoint splits raises whistle-to-hum top-1 from 0.631 (95% CI 0.581-0.681) to 0.673 (0.622-0.723), on known songs only. Third, two further generations of synthetic queries also fail to transfer: melody-conditioned whistles and la-la singing from a commercial neural music model leave real-whistle top-1 at 0.631, and an open-source pipeline (YingMusic-Singer-Plus singing with partly wrong, mumbled or nonsense lyrics, a DDSP whistle synthesizer, and a melody-error model fitted on real performers) gives at best 0.619 (paired difference -0.012, 95% CI -0.038 to +0.015). Across physics-based, commercial neural and open-source neural synthetic data, nothing transferred to real whistles, which points to collecting real queries as the missing step. Code is public; models and data will be released.

In medical imaging, the clinical value of Computed Tomography (CT) lies not only in depicting current disease status, but crucially in enabling longitudinal comparison of serial scans to determine disease evolution, a process that underpins response assessment, recurrence detection, and ongoing patient management. Yet, despite this central role of temporal comparison in clinical decision-making, existing medical foundation models remain largely confined to single-study understanding, leaving temporally grounded cross-examination insufficiently addressed. To address this gap, we study longitudinal imaging difference reporting, a task in which a model takes two temporally separated scans from the same patient and generates a clinically meaningful report describing interval changes between them. We introduce CT-ΔBench, a dedicated benchmark for this task with patient-level splitting to prevent information leakage. To better evaluate this task beyond surface-level text similarity, we further develop change-aware metrics specifically designed to capture clinically meaningful longitudinal changes, and conduct an independent physician validation to assess the reliability of the synthesized references and event extraction pipeline. We also compare direct paired-CT reasoning with an indirect two-stage pipeline that first generates single-timepoint reports and then performs textual differencing. Finally, we propose DeltaMed, a baseline model for direct paired-CT difference reporting, and train it on the benchmark training set. Together, these contributions lay the groundwork for temporally aware medical foundation models that better reflect real-world longitudinal clinical reasoning.

Large language models increasingly operate over large collections of tools, functions, APIs, and specialized agents. As the candidate action space grows, a function-calling model must process more schemas, consume more prompt tokens, and distinguish among increasingly similar or irrelevant alternatives. We study a complementary systems strategy: reduce the candidate set before language-model inference while leaving the downstream model unchanged. We introduce AgentWeave, a deterministic pre-inference routing layer that constructs a bounded model-visible action space using eligibility, requirement, capability, and routing signals. We evaluate AgentWeave with a frozen BFCL-derived routing-pressure protocol using the public MadeAgents/Hammer2.1-1.5b model. On 48 fresh BFCL V4 multiple-function tasks, AgentWeave achieves 6/48 (12.5%) native BFCL successes, whereas all-tools, deterministic random top-8, and semantic top-8 baselines each achieve 0/48. The paired success difference is +12.5 percentage points with a 10,000-resample paired bootstrap 95% confidence interval of +4.17 to +22.92 points and exact McNemar p=0.03125. Relative to all-tools exposure, AgentWeave presents 70.18% fewer tools, uses 61.70% fewer input tokens, and exhibits 50.95% lower mean local-model latency. The result is deliberately narrow: this is a BFCL-derived routing-pressure study rather than an official full BFCL leaderboard score, and absolute task success remains low. The evidence nevertheless shows that candidate-space construction can materially affect a fixed model's function-calling behavior and motivates evaluating routing as a distinct stage before model reasoning.

Large language model (LLM) agents solve complex tasks through multi-step interactions with external tools. These interactions often contain recurring local tool sequences. Treating such sequences as composite "Skills" can shorten tool-use trajectories and reduce repeated low-level decisions. However, when atomic tools and composite skills coexist, skill use becomes a policy problem: the agent must decide whether the current state requires atomic fine control or skill-level abstraction. In this paper, we argue that effective skill use should be studied as adaptive tool granularity selection. The most direct training signal for this problem is to compare the consequences of atomic and skill choices available from the same state. Based on this view, we propose CIPO, a Counterfactual Imagination Policy Optimization framework for adaptive tool granularity. CIPO constructs executable skills through budget-constrained mining of successful tool-use trajectories and trains granularity decisions with counterfactual branch rollouts. For each base rollout, CIPO branches at the first eligible granularity decision and replaces the chosen action with a feasible atomic or skill alternative. The paired outcome difference serves as a supplementary reward for policy optimization. Experiments across multiple benchmarks and model backbones show that CIPO improves task success and decision efficiency over baselines. Further analyses show that CIPO learns effective skill use by improving the choice between atomic tools and composite skills based on the current state, without simply increasing skill frequency.

Every production model is updated, by retraining, fine-tuning, quantization, or a silent vendor swap, and each update risks being worse than what it replaced. We formalize update promotion as certified paired risk-difference auditing. Our starting point is a support identity: the risk difference between two models lives on the inputs where they disagree, observable without labels. We build DISCERN, a sequential two-tier protocol. A zero-label tier certifies benign updates whose disagreement rate is below tolerance from unlabeled traffic alone. An audited tier labels only sampled disagreements through an anytime-valid confidence sequence, valid at every stopping time and under any label-routing rule, even an adversarial judge. We prove finite-sample validity and matching label-complexity bounds of order rho^2/eps^2 at the rate level, so exploiting free disagreement provably saves a factor 1/rho over any pairing-blind auditor, and the guarantee composes across an unbounded sequence of promotions from one error budget. Across 14,000+ replayed audit streams over 785 update pairs, including LoRA fine-tunes of language models up to 1.4B parameters, miscoverage is 0.0002 (nominal 5%), power 0.986 with zero false alarms, and 56% of benign updates certify with zero labels. Each audit emits a machine-checkable evidence record for post-market monitoring.

With the success of Neural Radiance Field (NeRF) in 3D-aware portrait editing, a variety of works have achieved promising results regarding both quality and 3D consistency. However, these methods heavily rely on per-prompt optimization when handling natural language as editing instructions. Due to the lack of labeled human face 3D datasets and effective architectures, the area of human-instructed 3D-aware editing for open-world portraits in an end-to-end manner remains under-explored. To solve this problem, we propose an end-to-end diffusion-based framework termed InstructPix2NeRF, which enables instructed 3D-aware portrait editing from a single open-world image with human instructions. At its core lies a conditional latent 3D diffusion process that lifts 2D editing to 3D space by learning the correlation between the paired images' difference and the instructions via triplet data. With the help of our proposed token position randomization strategy, we could even achieve multi-semantic editing through one single pass with the portrait identity well-preserved. Besides, we further propose an identity consistency module that directly modulates the extracted identity signals into our diffusion process, which increases the multi-view 3D identity consistency. Extensive experiments verify the effectiveness of our method and show its superiority against strong baselines quantitatively and qualitatively. Source code and pre-trained models can be found on our project page: https://mybabyyh.github.io/InstructPix2NeRF.

Population-level behavior in large-language-model (LLM) agents cannot be characterized by single-agent benchmarks. We introduce PV-SST, a peer-voted social-platform testbed, and report a separately frozen, preregistered matched-exposure experiment spanning four topics, four unused seeds, four open-weight model families, and three prespecified larger variants. The experiment comprises 448 trials and 112 complete model-by-topic-by-seed blocks. Relative to a topic-only control, a feed of previous-round peer posts ranked by peer-generated likes increases final-round lexical similarity in both the four-family core panel (paired mean difference +0.0082 TF-IDF cosine units, 95% block-bootstrap CI [0.0043, 0.0121], randomization p=0.000105, n=64 blocks) and the three-variant size extension (+0.0109 [0.0069, 0.0151], p=0.000001, n=48). This contrast bundles peer-post exposure with ranking and therefore does not identify a ranking-only effect. Opposite-side survival falls in the core panel (-3.9 percentage points [-6.8, -1.6], p=0.0068) but not conclusively in the larger variants (-1.0 pp [-3.1, 0.4], p=0.50). Holding adversarial impressions fixed, four distributed sources do not reliably move honest-agent stance more than one source. The preregistered distributed-minus-single contrast is positive but inconclusive in the core panel (+0.057 [-0.009, 0.125], p=0.112) and negative in the larger variants (-0.040 [-0.113, 0.035], p=0.332), failing the prespecified cross-model and cross-topic consistency criterion. Thus the robust result is lexical convergence under the tested peer-ranked feed, not general opinion capture or a general coordination advantage. The study evaluates synthetic LLM-agent populations; it does not estimate effects on people or production platforms.

A common small-model deployment runs one shared backbone with several LoRA specialists that answer over the same context. Serving them naively re-prefills that shared context once per specialist. We study a narrow, practical question: for already-trained standard LoRA adapters -- not adapters retrained for cache compatibility -- how much task quality is preserved if the backbone's prefill KV cache is computed once and reused across specialists, and what does that buy in serving cost? On a Qwen3-1.7B backbone with two adapters (extractive QA on HotpotQA, arithmetic reasoning on GSM8K), we sweep the boundary at which the specialist takes over from the reused base cache and measure paired quality differences and serving cost. Full-prefix reuse had the lowest prefill cost and a small quality difference on held-out GSM8K (Delta = -4.6 EM at a 160-token budget; -3.0 at 320 tokens; -0.8 under a second training seed -- all favoring native, only the first excluding zero, and the magnitude not consistent). Partial recomputation provided no demonstrated advantage. Neither quality equivalence nor a general boundary-selection rule is established. We also report a closed-form ridge KV translator that did not beat direct reuse, and specialist-dependence contrasts whose intervals all include zero. The measured serving benefit is warm-cache time-to-first-token, which grows with context (~16x at 8K); two-branch peak memory was only 12% lower and, on inspection, the prefix was never physically shared across branches -- this implementation reuses KV values but copies their storage, so shared-cache memory savings are not achieved.

Video-language models are increasingly used as judges of video understanding, both for evaluating model outputs and for training reward models. Whether their judgments remain reliable when the evidence is buried in day-long videos has yet to be established. Existing benchmarks cannot answer this. Their videos are typically only a few minutes long, many answer pairs can be separated from the transcript alone, and collecting human judgments does not scale to ultra-long videos. We introduce PlaylistEval, an agentic framework that builds video-language judge benchmarks over 100-hour playlist collection without human annotation. It automatically generates questions with paired answers whose differences are controlled by causal degradation, so that every pair demands retrieval across the collection. The resulting benchmark contains 630 pairs across seven domains spanning both static and dynamic knowledge, and on a stratified subset of 152 pairs it agrees with human judgments 93.0% of the time (IAA 0.781). Evaluating 17 omnimodal and multimodal models from eight families reveals that frontier judges reach only 75.4% pairwise accuracy, while open-source judge models perform far behind. We further show that both retrieval and final judgment depend on using multiple modalities, and that judge accuracy degrades as the playlist set grows. We release our pipeline, benchmark, and evaluation code at https://playlisteval.github.io.

Art reinterpretation is the practice of creating a variation of a reference work, making a paired artwork that exhibits a distinct artistic style. We ask if such an image pair can be used to customize a generative model to capture the demonstrated stylistic difference. We propose Pair Customization, a new customization method that learns stylistic difference from a single image pair and then applies the acquired style to the generation process. Unlike existing methods that learn to mimic a single concept from a collection of images, our method captures the stylistic difference between paired images. This allows us to apply a stylistic change without overfitting to the specific image content in the examples. To address this new task, we employ a joint optimization method that explicitly separates the style and content into distinct LoRA weight spaces. We optimize these style and content weights to reproduce the style and content images while encouraging their orthogonality. During inference, we modify the diffusion process via a new style guidance based on our learned weights. Both qualitative and quantitative experiments show that our method can effectively learn style while avoiding overfitting to image content, highlighting the potential of modeling such stylistic differences from a single image pair.

Motion instruction generation in cross-video comparison aims to produce corrective feedback that describes the differences between a query and a reference motion. However, existing models often generate instructions that exhibit motion hallucinations, failing to reflect actual kinematic differences between paired videos. To systematically investigate these hallucinations, we introduce MotionHalluc, a dedicated benchmark for evaluating motion hallucinations in paired-video comparison. MotionHalluc comprises 1540 fine-grained questions over 553 video pairs, evaluating hallucinations along three core dimensions: (1)directional hallucination, (2)attributional hallucination, and (3)temporal hallucination. Extensive evaluations of state-of-the-art large multimodal models demonstrate high susceptibility to these hallucinations. Furthermore, we provide Perceive-Parse-Verify (PPV) as a training-free measurements extraction and verification baseline that converts candidate instructions into executable measurement queries and supplies kinematic measurements at inference time. Our results show that this simple measurements injection yields an average 10.6% performance gain across models, suggesting that motion reasoning with explicit quantitative measurements is a key factor in reducing hallucinations in cross-video comparison. Our code and dataset will be made publicly available upon acceptance.

Distilling knowledge from human demonstrations is a promising way for robots to learn and act. Existing work often overlooks the differences between humans and robots, producing unsatisfactory results. In this paper, we study how perfectly aligned human-robot pairs benefit robot learning. Capitalizing on VR-based teleportation, we introduce H\&R, a third-person dataset with 2,600 episodes, each of which captures the fine-grained correspondence between human hand and robot gripper. Inspired by the recent success of diffusion models, we introduce Human2Robot, an end-to-end diffusion framework that formulates learning from human demonstration as a generative task. Human2Robot fully explores temporal dynamics in human videos to generate robot videos and predict actions at the same time. Through comprehensive evaluations of 4 carefully selected tasks in real-world settings, we demonstrate that Human2Robot can not only generate high-quality robot videos but also excels in seen tasks and generalizing to different positions, unseen appearances, novel instances, and even new backgrounds and task types.

Reinforcement learning (RL) has become a standard technique for post-training diffusion-based image synthesis models, as it enables learning from reward signals to explicitly improve desirable aspects such as image quality and prompt alignment. In this paper, we propose an online RL variant that reduces the variance in the model updates by sampling paired trajectories and pulling the flow velocity in the direction of the more favorable image. Unlike existing methods that treat each sampling step as a separate policy action, we consider the entire sampling process as a single action. We experiment with both high-quality vision language models and off-the-shelf quality metrics for rewards, and evaluate the outputs using a broad set of metrics. Our method converges faster and yields higher output quality and prompt alignment than previous approaches.

In this paper, we introduce the task of automatically generating text to describe the differences between two similar images. We collect a new dataset by crowd-sourcing difference descriptions for pairs of image frames extracted from video-surveillance footage. Annotators were asked to succinctly describe all the differences in a short paragraph. As a result, our novel dataset provides an opportunity to explore models that align language and vision, and capture visual salience. The dataset may also be a useful benchmark for coherent multi-sentence generation. We perform a firstpass visual analysis that exposes clusters of differing pixels as a proxy for object-level differences. We propose a model that captures visual salience by using a latent variable to align clusters of differing pixels with output sentences. We find that, for both single-sentence generation and as well as multi-sentence generation, the proposed model outperforms the models that use attention alone.

Example-based guidance is widely used to improve mathematical reasoning at inference time, yet its effectiveness is highly unstable across problems and models-even when the guidance is correct and problem-relevant. We show that this instability arises from a previously underexplored gap between strategy usage-whether a reasoning strategy appears in successful solutions-and strategy executability-whether the strategy remains effective when instantiated as guidance for a target model. Through a controlled analysis of paired human-written and model-generated solutions, we identify a systematic dissociation between usage and executability: human- and model-derived strategies differ in structured, domain-dependent ways, leading to complementary strengths and consistent source-dependent reversals under guidance. Building on this diagnosis, we propose Selective Strategy Retrieval (SSR), a test-time framework that explicitly models executability by selectively retrieving and combining strategies using empirical, multi-route, source-aware signals. Across multiple mathematical reasoning benchmarks, SSR yields reliable and consistent improvements over direct solving, in-context learning, and single-source guidance, improving accuracy by up to +13 points on AIME25 and +5 points on Apex for compact reasoning models. Code and benchmark are publicly available at: https://github.com/lwd17/strategy-execute-pipeline.

How do two sets of images differ? Discerning set-level differences is crucial for understanding model behaviors and analyzing datasets, yet manually sifting through thousands of images is impractical. To aid in this discovery process, we explore the task of automatically describing the differences between two sets of images, which we term Set Difference Captioning. This task takes in image sets D_A and D_B, and outputs a description that is more often true on D_A than D_B. We outline a two-stage approach that first proposes candidate difference descriptions from image sets and then re-ranks the candidates by checking how well they can differentiate the two sets. We introduce VisDiff, which first captions the images and prompts a language model to propose candidate descriptions, then re-ranks these descriptions using CLIP. To evaluate VisDiff, we collect VisDiffBench, a dataset with 187 paired image sets with ground truth difference descriptions. We apply VisDiff to various domains, such as comparing datasets (e.g., ImageNet vs. ImageNetV2), comparing classification models (e.g., zero-shot CLIP vs. supervised ResNet), summarizing model failure modes (supervised ResNet), characterizing differences between generative models (e.g., StableDiffusionV1 and V2), and discovering what makes images memorable. Using VisDiff, we are able to find interesting and previously unknown differences in datasets and models, demonstrating its utility in revealing nuanced insights.

Robustness evaluation must examine diverse visual perturbations, while benchmarks cover only some real-world conditions and physical testing is costly. Paired evaluations link clean and perturbed predictions for the same image, capturing changes in correctness, confidence, and acceptance beyond aggregate accuracy. We investigate how this image correspondence supports two needs in robustness evaluation: interpreting paired evaluation results and prioritizing samples for physical testing. To interpret paired evaluation results, we fix both sets of prediction records and vary their correspondence within each class. We prove that classwise correct-correct counts give the same sharp bounds on lost acceptance and mean true-class probability decrease among retained-correct inputs as any feasible five-state refinement. Distinguishing persistent from changed wrong answers can further constrain accepted-error transitions, while shared correspondence can establish policy orderings left unresolved by separate cost intervals. To prioritize samples for physical testing, we retain each image's synthetic responses and rank clean-correct images by their mean true-class probability under corruption. Across 44 classifiers, testing the highest-risk 20% finds 67% and 45% of failures under mild screen and print recaptures, versus 58% and 36% for clean confidence and 60% and 37% for an equal-size natural-transformation average. With both probability averaging and an A3Rank scoring adaptation, the tested corruption set yields higher mean failure recall than the natural-transform set; differences between scores depend on the source and budget. Together, these findings show that the value of correspondence depends on the evaluation objective: classwise counts suffice for specified reliability bounds, while image-specific synthetic responses improve the allocation of physical tests within the evaluated pool.

Automated shape repair approaches currently lack access to datasets that describe real-world damaged geometry. We present Fantastic Breaks (and Where to Find Them: https://terascale-all-sensing-research-studio.github.io/FantasticBreaks), a dataset containing scanned, waterproofed, and cleaned 3D meshes for 150 broken objects, paired and geometrically aligned with complete counterparts. Fantastic Breaks contains class and material labels, proxy repair parts that join to broken meshes to generate complete meshes, and manually annotated fracture boundaries. Through a detailed analysis of fracture geometry, we reveal differences between Fantastic Breaks and synthetic fracture datasets generated using geometric and physics-based methods. We show experimental shape repair evaluation with Fantastic Breaks using multiple learning-based approaches pre-trained with synthetic datasets and re-trained with subset of Fantastic Breaks.

This paper argues for the widest possible use of bootstrap confidence intervals for comparing NLP system performances instead of the state-of-the-art status (SOTA) and statistical significance testing. Their main benefits are to draw attention to the difference in performance between two systems and to help assessing the degree of superiority of one system over another. Two cases studies, one comparing several systems and the other based on a K-fold cross-validation procedure, illustrate these benefits. A python module for obtaining these confidence intervals as well as a second function implementing the Fisher-Pitman test for paired samples are freely available on PyPi.

In-context learning (ICL) enables models to infer tasks from demonstrations, but existing benchmarks generally lack matched text and image versions needed to compare ICL performance across modalities. We introduce TwinICL, a procedurally generated benchmark providing such pairs for controlled comparison. Across six open-weight models and 38 tasks, multimodal ICL consistently underperforms text-only ICL, with gaps varying by task family. To test whether this gap can be recovered, we target visual access, task framing, and reasoning through three interventions. Their combination recovers strong multimodal ICL performance on a diagnostic subset, despite limited or inconsistent individual effects. To distinguish difficulties in executing tasks from those in inferring them, we evaluate models with explicit task instructions, revealing a modality gap even when the task is known. We then examine how adding demonstration inputs and outputs reshapes this gap, highlighting demonstrations' dual role as additional context to process and evidence about the task. The dataset is available at https://github.com/lab-flair/TwinICL.

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.

This study examines how persona prompting shapes language generated by two multimodal large language models in urban perception, a setting for examining subjective interpretations of shared visual evidence. We organize outputs into three functional layers: descriptive grounding (captions), intermediate semantic layer (perception tags), and interpretive framing (justifications). Using approximately 60,000 persona-conditioned annotations from each of two MLLMs, Qwen3-VL and Gemma4, we find that captions converge strongly across persona profiles and show only small attribute-associated differences. Justifications vary substantially more: economic status produces the largest difference in both models, with political orientation and personality also prominent. Paired image-level comparisons confirm larger justification than caption differences for these three attributes. For perception tags, personas sharing the same attribute level produce more similar tag sets than personas with different attribute levels, with the largest separation observed for economic status. Exploratory topic analysis further suggests persona-specific evaluative emphasis. Across models, profile-pair similarity patterns are strongly correlated for all three output types, although agreement is lowest for justifications. Overall, persona prompting affects interpretive framing more strongly than descriptive grounding.

Text-to-audio-video (T2AV) generation is central to applications such as filmmaking and world modeling. However, current models often fail to produce physically plausible sounds. Previous benchmarks primarily focus on audio-video temporal synchronization, while largely overlooking explicit evaluation of audio-physics grounding, thereby limiting the study of physically plausible audio-visual generation. To address this issue, we present PhyAVBench, the first benchmark that systematically evaluates the audio-physics grounding capabilities of T2AV, image-to-audio-video (I2AV), and video-to-audio (V2A) models. PhyAVBench offers PhyAV-Sound-11K, a new dataset of 25.5 hours of 11,605 audible videos collected from 184 participants to ensure diversity and avoid data leakage. It contains 337 paired-prompt groups with controlled physical variations that drive sound differences, each grounded with an average of 17 videos and spanning 6 audio-physics dimensions and 41 fine-grained test points. Each prompt pair is annotated with the physical factors underlying their acoustic differences. Importantly, PhyAVBench leverages paired text prompts to evaluate this capability. We term this evaluation paradigm the Audio-Physics Sensitivity Test (APST) and introduce a novel metric, the Contrastive Physical Response Score (CPRS), which quantifies the acoustic consistency between generated videos and their real-world counterparts. We conduct a comprehensive evaluation of 17 state-of-the-art models. Our results reveal that even leading commercial models struggle with fundamental audio-physical phenomena, exposing a critical gap beyond audio-visual synchronization and pointing to future research directions. We hope PhyAVBench will serve as a foundation for advancing physically grounded audio-visual generation. Prompts, ground-truth, and generated video samples are available at https://github.com/imxtx/PhyAVBench.

Evaluating explainability techniques, such as SHAP and LRP, in the context of audio deepfake detection is challenging due to lack of clear ground truth annotations. In the cases when we are able to obtain the ground truth, we find that these methods struggle to provide accurate explanations. In this work, we propose a novel data-driven approach to identify artifact regions in deepfake audio. We consider paired real and vocoded audio, and use the difference in time-frequency representation as the ground-truth explanation. The difference signal then serves as a supervision to train a diffusion model to expose the deepfake artifacts in a given vocoded audio. Experimental results on the VocV4 and LibriSeVoc datasets demonstrate that our method outperforms traditional explainability techniques, both qualitatively and quantitatively.

Public leaderboards for coding agents typically rank systems by model name and pass rate, while the surrounding harness (the scaffold that issues tools, manages context, and decides when to stop) is often under-specified. Model-to-model comparison is valid when the harness is fixed; when it varies, performance and efficiency conflate model and scaffold effects. We evaluate Qwen 3.6 Plus and MiniMax M2.5 across three open-source harnesses (Goose, OpenCode, OpenHands-SDK) on a stratified 50-task subset of Terminal-Bench Pro. Harness choice induces up to a 40x difference in tokens per solved task, while paired within-model pass-rate differences remain 0-8 percentage points (95% paired-task bootstrap CIs include zero except for the largest gap). Failure fingerprints replicate across models (REASON for Goose, VERIFY/MAX_TURNS for OpenHands-SDK, idle-loop/TIME for OpenCode), indicating harness-level biases that are largely model-independent. For human-centered coding-agent evaluation, model name alone is an incomplete comparison unit: harness-model pairs determine real-world cost, latency, and oversight burden; no-action turns are a per-task wait tax, not just a token tax. We therefore recommend selecting harness-model pairs by pass rate under token/latency budgets, and reporting token usage, latency, and full harness specifications alongside any model comparison. We release anonymized configs, raw trial logs, aggregated snapshots, and analysis scripts.

We present SpaceTimePilot, a video diffusion model that disentangles space and time for controllable generative rendering. Given a monocular video, SpaceTimePilot can independently alter the camera viewpoint and the motion sequence within the generative process, re-rendering the scene for continuous and arbitrary exploration across space and time. To achieve this, we introduce an effective animation time-embedding mechanism in the diffusion process, allowing explicit control of the output video's motion sequence with respect to that of the source video. As no datasets provide paired videos of the same dynamic scene with continuous temporal variations, we propose a simple yet effective temporal-warping training scheme that repurposes existing multi-view datasets to mimic temporal differences. This strategy effectively supervises the model to learn temporal control and achieve robust space-time disentanglement. To further enhance the precision of dual control, we introduce two additional components: an improved camera-conditioning mechanism that allows altering the camera from the first frame, and CamxTime, the first synthetic space-and-time full-coverage rendering dataset that provides fully free space-time video trajectories within a scene. Joint training on the temporal-warping scheme and the CamxTime dataset yields more precise temporal control. We evaluate SpaceTimePilot on both real-world and synthetic data, demonstrating clear space-time disentanglement and strong results compared to prior work. Project page: https://zheninghuang.github.io/Space-Time-Pilot/ Code: https://github.com/ZheningHuang/spacetimepilot

Ensuring reliable model performance across diverse domains is a critical challenge in computational pathology. A particular source of variability in Whole-Slide Images is introduced by differences in digital scanners, thus calling for better scanner generalization. This is critical for the real-world adoption of computational pathology, where the scanning devices may differ per institution or hospital, and the model should not be dependent on scanner-induced details, which can ultimately affect the patient's diagnosis and treatment planning. However, past efforts have primarily focused on standard domain generalization settings, evaluating on unseen scanners during training, without directly evaluating consistency across scanners for the same tissue. To overcome this limitation, we introduce SCORPION, a new dataset explicitly designed to evaluate model reliability under scanner variability. SCORPION includes 480 tissue samples, each scanned with 5 scanners, yielding 2,400 spatially aligned patches. This scanner-paired design allows for the isolation of scanner-induced variability, enabling a rigorous evaluation of model consistency while controlling for differences in tissue composition. Furthermore, we propose SimCons, a flexible framework that combines augmentation-based domain generalization techniques with a consistency loss to explicitly address scanner generalization. We empirically show that SimCons improves model consistency on varying scanners without compromising task-specific performance. By releasing the SCORPION dataset and proposing SimCons, we provide the research community with a crucial resource for evaluating and improving model consistency across diverse scanners, setting a new standard for reliability testing.

Memory self-evolution uses task feedback to iteratively improve executable memory programs that store and retrieve information from past interactions. Existing approaches typically adopt holistic evolution, deriving revision directions from mixed feedback and judging progress by overall performance. This can obscure optimization directions and hide capability-specific gains offset by regressions elsewhere, leaving promising directions underexplored. We introduce capability-driven evolution, which extends search guidance from overall performance to individual capability dimensions, preserving promising revisions and expanding exploration beyond the boundaries of holistic evolution. We propose PrisMem, which uses dependency-aware capability selection to prioritize targets with potential cross-capability benefits and history-guided diagnosis to refine capability specialists. Trace-guided integration compares evaluated programs on paired differential cases, using their behavioral differences to consolidate complementary gains into a unified memory program. Experiments show that PrisMem outperforms the strongest baselines by 10.54 and 7.83 percentage points on BEAM-1M and LongMemEval-M, respectively, demonstrating its effectiveness on million-token histories.

The application of diffusion models in 3D LiDAR scene completion is limited due to diffusion's slow sampling speed. Score distillation accelerates diffusion sampling but with performance degradation, while post-training with direct policy optimization (DPO) boosts performance using preference data. This paper proposes Distillation-DPO, a novel diffusion distillation framework for LiDAR scene completion with preference aligment. First, the student model generates paired completion scenes with different initial noises. Second, using LiDAR scene evaluation metrics as preference, we construct winning and losing sample pairs. Such construction is reasonable, since most LiDAR scene metrics are informative but non-differentiable to be optimized directly. Third, Distillation-DPO optimizes the student model by exploiting the difference in score functions between the teacher and student models on the paired completion scenes. Such procedure is repeated until convergence. Extensive experiments demonstrate that, compared to state-of-the-art LiDAR scene completion diffusion models, Distillation-DPO achieves higher-quality scene completion while accelerating the completion speed by more than 5-fold. Our method is the first to explore adopting preference learning in distillation to the best of our knowledge and provide insights into preference-aligned distillation. Our code is public available on https://github.com/happyw1nd/DistillationDPO.

The ability to distinguish subtle differences between visually similar images is essential for diverse domains such as industrial anomaly detection, medical imaging, and aerial surveillance. While comparative reasoning benchmarks for vision-language models (VLMs) have recently emerged, they primarily focus on images with large, salient differences and fail to capture the nuanced reasoning required for real-world applications. In this work, we introduce VLM-SubtleBench, a benchmark designed to evaluate VLMs on subtle comparative reasoning. Our benchmark covers ten difference types - Attribute, State, Emotion, Temporal, Spatial, Existence, Quantity, Quality, Viewpoint, and Action - and curate paired question-image sets reflecting these fine-grained variations. Unlike prior benchmarks restricted to natural image datasets, our benchmark spans diverse domains, including industrial, aerial, and medical imagery. Through extensive evaluation of both proprietary and open-source VLMs, we reveal systematic gaps between model and human performance across difference types and domains, and provide controlled analyses highlighting where VLMs' reasoning sharply deteriorates. Together, our benchmark and findings establish a foundation for advancing VLMs toward human-level comparative reasoning.

Recent studies extend the autoregression paradigm to text-to-image generation, achieving performance comparable to diffusion models. However, our new PairComp benchmark -- featuring test cases of paired prompts with similar syntax but different fine-grained semantics -- reveals that existing models struggle with fine-grained text-image alignment thus failing to realize precise control over visual tokens. To address this, we propose FocusDiff, which enhances fine-grained text-image semantic alignment by focusing on subtle differences between similar text-image pairs. We construct a new dataset of paired texts and images with similar overall expressions but distinct local semantics, further introducing a novel reinforcement learning algorithm to emphasize such fine-grained semantic differences for desired image generation. Our approach achieves state-of-the-art performance on existing text-to-image benchmarks and significantly outperforms prior methods on PairComp.

Predicting transcriptional responses to genetic perturbations is central to understanding gene function. Existing predictors primarily rely on transcriptomic measurements, although chromatin accessibility provides complementary information about the cellular context in which perturbations act. Using this information requires linking chromatin context to specific perturbations and accounting for baseline differences between independently sampled control and perturbed populations. We propose ChromaPert, a chromatin-informed framework for predicting state-dependent perturbation responses. ChromaPert combines molecular and DNA-informed locus priors with paired control RNA--ATAC features to jointly represent target identity and measured cellular context. Its Chromatin-Guided Response Router (CGRR) builds a source bank of control-derived expression offsets and retrieves relevant sources using perturbation/locus or control-ATAC similarity. These offsets account for baseline differences, while conditional transport flow learns the remaining response. Across three unseen-target and held-out-state settings on K562 CAT-ATAC and Perturb-Multiome, ChromaPert achieves the highest all-gene response correlation among evaluated methods. When transferring known perturbations to held-out states, it improves correlation by 23.2% across all genes and 46.2% for the twenty most responsive genes over the strongest respective baselines. Correct RNA--ATAC pairing improves recovery of chromatin-associated response slopes over shuffled pairing, while cross-state predictions retain state-specific responses to the same perturbation.

Accurate estimation of pasture biomass is important for decision-making in livestock production systems. Estimates of pasture biomass can be used to manage stocking rates to maximise pasture utilisation, while minimising the risk of overgrazing and promoting overall system health. We present a comprehensive dataset of 1,162 annotated top-view images of pastures collected across 19 locations in Australia. The images were taken across multiple seasons and include a range of temperate pasture species. Each image captures a 70cm * 30cm quadrat and is paired with on-ground measurements including biomass sorted by component (green, dead, and legume fraction), vegetation height, and Normalized Difference Vegetation Index (NDVI) from Active Optical Sensors (AOS). The multidimensional nature of the data, which combines visual, spectral, and structural information, opens up new possibilities for advancing the use of precision grazing management. The dataset is released and hosted in a Kaggle competition that challenges the international Machine Learning community with the task of pasture biomass estimation. The dataset is available on the official Kaggle webpage: https://www.kaggle.com/competitions/csiro-biomass

The rise of large language models (LLMs) has created a significant disparity: industrial research labs with their computational resources, expert teams, and advanced infrastructures, can effectively fine-tune LLMs, while individual developers and small organizations face barriers due to limited resources. In this paper, we aim to bridge this gap by presenting a comprehensive study on supervised fine-tuning of LLMs using instruction-tuning datasets spanning diverse knowledge domains and skills. We focus on small-sized LLMs (3B to 7B parameters) for their cost-efficiency and accessibility. We explore various training configurations and strategies across four open-source pre-trained models. We provide detailed documentation of these configurations, revealing findings that challenge several common training practices, including hyperparameter recommendations from TULU and phased training recommended by Orca. Key insights from our work include: (i) larger batch sizes paired with lower learning rates lead to improved model performance on benchmarks such as MMLU, MTBench, and Open LLM Leaderboard; (ii) early-stage training dynamics, such as lower gradient norms and higher loss values, are strong indicators of better final model performance, enabling early termination of sub-optimal runs and significant computational savings; (iii) through a thorough exploration of hyperparameters like warmup steps and learning rate schedules, we provide guidance for practitioners and find that certain simplifications do not compromise performance; and (iv) we observed no significant difference in performance between phased and stacked training strategies, but stacked training is simpler and more sample efficient. With these findings holding robustly across datasets and models, we hope this study serves as a guide for practitioners fine-tuning small LLMs and promotes a more inclusive environment for LLM research.

Large language model (LLM) agents increasingly take actions (screening applicants, recommending credit, triaging patients), yet fairness for LLMs is still measured by grading answers. We introduce AgentFairBench, a cheap, reproducible, multi-domain benchmark for demographic disparity in the actions of LLM agents. Grounded in a companion framework, the Bias Conduction Framework (BCF, restated here), it spans three regulator-anchored domains: hiring, lending, and medical triage. Synthetic, demographic-neutral profiles are evaluated in counterfactual matched sets that vary only a name-coded race x gender signal (in the Bertrand Mullainathan tradition), under four agent scaffolds of increasing agency (direct, chain-of-thought, multi-agent deliberation, tool-augmented). A NumPy-only harness computes counterfactual flip rate, mean absolute score difference (MASD), action-rate disparity, and tool-invocation disparity, with bootstrap confidence intervals, paired tests, and false-discovery-rate control, for single-digit dollars per model. A live leaderboard with a held-out private split and a contamination canary admits external models by submission. Our pilot (864 decisions plus a test-retest replication) carries a methodological lesson: comparing a six-group score spread against a two-run noise difference overstates disparity by ~ 2.4X through statistic arity alone. Against an arity matched noise floor and an omnibus group test, claude haiku 4 5 shows no demographic effect above sampling noise (0 of 120 pairwise and 0 of 9 omnibus contrasts survive correction); a planted-bias test confirms the instrument detects disparity when present. The contribution is a sound, sensitive, adoption-ready instrument, the arity matched null methodology, and open artifacts to scale it. Code, data, and harness are released under open licenses, with an anonymized review artifact.

Representing wild sounds as images is an important but challenging task due to the lack of paired datasets between sound and images and the significant differences in the characteristics of these two modalities. Previous studies have focused on generating images from sound in limited categories or music. In this paper, we propose a novel approach to generate images from in-the-wild sounds. First, we convert sound into text using audio captioning. Second, we propose audio attention and sentence attention to represent the rich characteristics of sound and visualize the sound. Lastly, we propose a direct sound optimization with CLIPscore and AudioCLIP and generate images with a diffusion-based model. In experiments, it shows that our model is able to generate high quality images from wild sounds and outperforms baselines in both quantitative and qualitative evaluations on wild audio datasets.

Real-time full-duplex speech models can listen while speaking, enabling natural interaction without rigid turn boundaries. Existing benchmarks evaluate turn-taking, interruption handling and multi-round dialogue, but largely centre on a designated user rather than an assistant participating in a shared conversation among several people. We introduce Duplex-MPE to evaluate when such an assistant should answer, remain silent or stop speaking. The benchmark contains 2,000 scenarios with three or four human speakers and one assistant, each paired across explicit and implicit addressing of the same request. Models receive continuous conversation audio without transcripts or supplied turn boundaries. Four scores measure fresh response initiation, answer accuracy, silence preservation and stopping when a human resolves a request. We evaluate five open-weight speech systems: MiniCPM-o 4.5, Moshi, FLM-Audio, Voila and Freeze-Omni. MiniCPM-o 4.5 leads on three scored capabilities, while frequent speech from other systems can coexist with inaccurate answers or failures to remain silent. A transcript-based Gemini 3.1 Pro reference responds 64.3 percentage points more often to explicit than implicit requests; paired tests detect no significant response-rate difference for the speech systems.

We introduce MuirBench, a comprehensive benchmark that focuses on robust multi-image understanding capabilities of multimodal LLMs. MuirBench consists of 12 diverse multi-image tasks (e.g., scene understanding, ordering) that involve 10 categories of multi-image relations (e.g., multiview, temporal relations). Comprising 11,264 images and 2,600 multiple-choice questions, MuirBench is created in a pairwise manner, where each standard instance is paired with an unanswerable variant that has minimal semantic differences, in order for a reliable assessment. Evaluated upon 20 recent multi-modal LLMs, our results reveal that even the best-performing models like GPT-4o and Gemini Pro find it challenging to solve MuirBench, achieving 68.0% and 49.3% in accuracy. Open-source multimodal LLMs trained on single images can hardly generalize to multi-image questions, hovering below 33.3% in accuracy. These results highlight the importance of MuirBench in encouraging the community to develop multimodal LLMs that can look beyond a single image, suggesting potential pathways for future improvements.

Aligning Text-to-Image (T2I) generation models with human preferences increasingly relies on image reward models that score or rank generated images according to prompt alignment and perceptual quality. Existing reward models are commonly trained as Bradley-Terry (BT) preference models on large-scale human preference corpora, making them costly to train, difficult to adapt, and opaque in their evaluation criteria. Meanwhile, Vision-Language Model (VLM) judges can provide more fine-grained assessments through textual rubrics, but their manually designed or heuristically generated scoring rules may fail to reliably reflect human preferences. In this paper, we propose AutoRubric-T2I, the first rubric learning framework in T2I that automatically synthesizes and selects explicit rubrics for guiding VLM judges. AutoRubric-T2I first synthesizes reasoning traces from preference pairs into candidate rubrics, then uses a VLM judge to score paired images under each rubric, producing pairwise rubric-score differences for preference learning. To remove noisy and redundant rules, we further employ a ell_1-Regularized Logistic Regression Refiner, which selects the Top-N most discriminative rubrics. Extensive evaluations show that AutoRubric-T2I produces high-quality, interpretable reward signals using less than 0.01% of the annotated preference data, substantially reducing the need for large-scale reward-model training. On image reward benchmarks such as MMRB2, AutoRubric-T2I outperforms strong reward model baselines. We further validate AutoRubric-T2I as an RL reward on downstream T2I tasks, including TIIF and UniGenBench++, where it improves generation quality over scalar reward models using the Flow-GRPO pipeline on diffusion models.

We describe our submission to the MedReason 2026 challenge, covering multiple-choice (MCQ) and open-ended (OE) medical visual question answering (VQA) under fully offline, containerized inference. Our first finding is that MCQ retrieval must compare answer semantics rather than answer labels: labels are independently assigned per question, so copying a retrieved neighbor's label transfers no useful information, whereas scoring each current option's text against correct-answer text from similar training cases raises retrieval-only accuracy from 20.0\% to 57.5\% on a 200-case retrieval-excluded development holdout. Our second finding attributes the submitted system's accuracy: holding the task-specific MCQ Low-Rank Adaptation (LoRA) adapter fixed and varying the number \(k\) of in-prompt retrieved examples changes accuracy by at most one case --- 187/200 (93.5\%) at both \(k=0\) and the adapter's training-time \(k=1\), 188/200 (94.0\%) at the packaged runtime's default \(k=3\) --- and the submitted confidence-gated override adds no net accuracy on top of \(k=3\), selecting the VLM in 198/200 cases. With the final MCQ adapter fixed, retrieval changes accuracy by at most one case, and gating provides no net gain. On 20 OE cases, token-F1 and RaTEScore~zhao2024ratescore decrease as \(k\) grows, but paired sign tests on token-F1 differences are nonsignificant (\(p \ge 0.29\)); a single-annotator comparison found 6/20 wrong-anchor errors for the final configuration and 14/20 for an earlier configuration that jointly differed in routing, adapter, and prompting. The system reaches 94.0\% MCQ accuracy on the development holdout and 93.20\% on the organizer's official pre-evaluation, versus 29.43\% for the off-the-shelf reference baseline, while both of the organizer's open-ended scores are lower than that baseline's (ground-truth agreement 1.245 versus 1.588, visual accuracy 1.995 versus 2.696, each out of 4).

A benchmark score compares a system output against a reference, and methodological attention falls almost entirely on the first term. We measure the second. The retained annotation record of a six-language benchmark for personally identifiable information contains two independent annotator labellings, the aggregate shipped as gold, and a reviewer gold from independent expert re-annotation of a sample. Using it, we hold the scored output fixed and exchange the reference. The score moves by 4.95 F1 points for one output and 2.00 for the other (95% CIs [3.23, 6.27] and [0.26, 3.37]), and by 7.55 in the worst language. The comparison between two outputs moves as well: the paired interaction between reference and system is +2.95 points (CI [+1.97, +3.87]), survives correction for multiple testing, and changes one language's margin outright. The cause is an undocumented aggregation default that usually kept one annotator when adjudication did not fire, making the shipped reference a partial copy of an output being scored. We report the resulting reference-sensitivity band, show how to compute one from any retained annotation record, and argue that the quantity belongs beside the score.

In contrastive learning, two views of an original image, generated by different augmentations, are considered a positive pair, and their similarity is required to be high. Similarly, two views of distinct images form a negative pair, with encouraged low similarity. Typically, a single similarity measure, provided by a lone projection head, evaluates positive and negative sample pairs. However, due to diverse augmentation strategies and varying intra-sample similarity, views from the same image may not always be similar. Additionally, owing to inter-sample similarity, views from different images may be more akin than those from the same image. Consequently, enforcing high similarity for positive pairs and low similarity for negative pairs may be unattainable, and in some cases, such enforcement could detrimentally impact performance. To address this challenge, we propose using multiple projection heads, each producing a distinct set of features. Our pre-training loss function emerges from a solution to the maximum likelihood estimation over head-wise posterior distributions of positive samples given observations. This loss incorporates the similarity measure over positive and negative pairs, each re-weighted by an individual adaptive temperature, regulated to prevent ill solutions. Our approach, Adaptive Multi-Head Contrastive Learning (AMCL), can be applied to and experimentally enhances several popular contrastive learning methods such as SimCLR, MoCo, and Barlow Twins. The improvement remains consistent across various backbones and linear probing epochs, and becomes more significant when employing multiple augmentation methods.

How do two individuals differ when performing the same action? In this work, we introduce Video Action Differencing (VidDiff), the novel task of identifying subtle differences between videos of the same action, which has many applications, such as coaching and skill learning. To enable development on this new task, we first create VidDiffBench, a benchmark dataset containing 549 video pairs, with human annotations of 4,469 fine-grained action differences and 2,075 localization timestamps indicating where these differences occur. Our experiments demonstrate that VidDiffBench poses a significant challenge for state-of-the-art large multimodal models (LMMs), such as GPT-4o and Qwen2-VL. By analyzing failure cases of LMMs on VidDiffBench, we highlight two key challenges for this task: localizing relevant sub-actions over two videos and fine-grained frame comparison. To overcome these, we propose the VidDiff method, an agentic workflow that breaks the task into three stages: action difference proposal, keyframe localization, and frame differencing, each stage utilizing specialized foundation models. To encourage future research in this new task, we release the benchmark at https://huggingface.co/datasets/jmhb/VidDiffBench and code at http://jmhb0.github.io/viddiff.

Pairwise re-ranking models predict which of two documents is more relevant to a query and then aggregate a final ranking from such preferences. This is often more effective than pointwise re-ranking models that directly predict a relevance value for each document. However, the high inference overhead of pairwise models limits their practical application: usually, for a set of k documents to be re-ranked, preferences for all k^2-k comparison pairs excluding self-comparisons are aggregated. We investigate whether the efficiency of pairwise re-ranking can be improved by sampling from all pairs. In an exploratory study, we evaluate three sampling methods and five preference aggregation methods. The best combination allows for an order of magnitude fewer comparisons at an acceptable loss of retrieval effectiveness, while competitive effectiveness is already achieved with about one third of the comparisons.

Ambient AI scribes draft clinical notes, and published audits find their dominant error is omission: information the encounter established that the note fails to record. The standard check is an LLM judge: a second model reads the note against the transcript and flags problems. We ask whether judges detect omissions. Public corpora cannot supply the answer key: their clinician reference notes and transcripts are materially discrepant. Our benchmark has 500 single-error note pairs from audited fact sheets, 298 with a named fact certainly absent and 202 added-or-altered controls. Across eight judge designs, paired discrimination (the flawed note below its clean twin, 0.5 a coin flip) reads 0.79-0.94 on added or altered content and 0.50-0.63 on omissions. On single notes, no design flags omissions reliably more often than perfect notes. Wording changes, voting and GEPA prompt optimisation move the operating point without creating usable detection. Restructuring the task recovers it: list the facts the transcript establishes, then check the note for each. Two methods reach it independently and trade off: a per-fact pipeline, and a GEPA-evolved prompt doing the same in one call. The pipeline's flags name the missing fact and its severity at 2.7% false alarms. The single call detects more (36.9% against 24.6%, p=0.002) at 6.2% false alarms and a tenth of the cost per note. A physician author validated 70 items and, where the two routes disagree, sided with the pipeline on 10 of 10 (p=0.002). A second clinician, not an author, graded the severity rubric blind and agrees to within a grade. On real vendor notes from a companion census no benchmark threshold transfers, but the re-calibrated single call detects more than the best of the eight at half its false-alarm rate. Omissions whose fact is restated elsewhere defeat both routes. We release the benchmark, prompts and judgements.

Leaderboards rank long-horizon agents by their final outputs. Yet a higher score alone does not establish whether two systems are comparable or which stage accounts for the difference. Unequal evidence, inputs, or budgets can affect scores, and statistical corrections do not remove this mismatch. We introduce compositional controllability to address these questions. A comparison window covers one stage, several stages, or the whole agent. Our central result bounds the gap between observed and controlled score differences using only nuisance outside the window. This yields an admissibility test applied before scores are inspected. Inadmissible comparisons are refused. For admissible pairs, an ordering is certified only when the score gap exceeds the combined sampling and nuisance radii; otherwise, it remains undecided. These decisions give each system a rank interval. We introduce BioLitBench, a benchmark of 2,042 biomedical articles represented as structured claim graphs. Among seven published pipelines, a conventional statistical analysis declares a winner in 14 of 21 pairwise comparisons. Yet the top-ranked system alone received the target review's bibliography. To isolate pipeline performance, our test requires matched inputs and a fixed backbone model. It refuses 11 of the 21 comparisons, including every comparison involving the top-ranked system. Seven of the 14 conventional conclusions fall within these refused pairs. The same comparison windows support stage-level training. We train SCRIBE on Qwen3.8-27B using rewards measured at each stage's exit. Under matched evidence, SCRIBE achieves a certified rank interval of [1,2], with certified advantages over all evaluated published pipelines and the evaluated Claude and OpenAI agents. Under same pool, SCRIBE matches the strongest published retriever and is certified above three published pipelines.

Shepard's universal law of generalization is a remarkable hypothesis about how intelligent organisms should perceive similarity. In its broadest form, the universal law states that the level of perceived similarity between a pair of stimuli should decay as a concave function of their distance when embedded in an appropriate psychological space. While extensively studied, evidence in support of the universal law has relied on low-dimensional stimuli and small stimulus sets that are very different from their real-world counterparts. This is largely because pairwise comparisons -- as required for similarity judgments -- scale quadratically in the number of stimuli. We provide direct evidence for the universal law in a naturalistic high-dimensional regime by analyzing an existing dataset of 214,200 human similarity judgments and a newly collected dataset of 390,819 human generalization judgments (N=2406 US participants) across three sets of natural images.

Improvements in language models are often driven by improving the quality of the data we train them on, which can be limiting when strong supervision is scarce. In this work, we show that paired preference data consisting of individually weak data points can enable gains beyond the strength of each individual data point. We formulate the delta learning hypothesis to explain this phenomenon, positing that the relative quality delta between points suffices to drive learning via preference tuning--even when supervised finetuning on the weak data hurts. We validate our hypothesis in controlled experiments and at scale, where we post-train 8B models on preference data generated by pairing a small 3B model's responses with outputs from an even smaller 1.5B model to create a meaningful delta. Strikingly, on a standard 11-benchmark evaluation suite (MATH, MMLU, etc.), our simple recipe matches the performance of Tulu 3, a state-of-the-art open model tuned from the same base model while relying on much stronger supervisors (e.g., GPT-4o). Thus, delta learning enables simpler and cheaper open recipes for state-of-the-art post-training. To better understand delta learning, we prove in logistic regression that the performance gap between two weak teacher models provides useful signal for improving a stronger student. Overall, our work shows that models can learn surprisingly well from paired data that might typically be considered weak.

Determining sentence pair similarity is crucial for various NLP tasks. A common technique to address this is typically evaluated on a continuous semantic textual similarity scale from 0 to 5. However, based on a linguistic observation in STS annotation guidelines, we found that the score in the range [4,5] indicates an upper-range sample, while the rest are lower-range samples. This necessitates a new approach to treating the upper-range and lower-range classes separately. In this paper, we introduce a novel embedding space decomposition method called MixSP utilizing a Mixture of Specialized Projectors, designed to distinguish and rank upper-range and lower-range samples accurately. The experimental results demonstrate that MixSP decreased the overlap representation between upper-range and lower-range classes significantly while outperforming competitors on STS and zero-shot benchmarks.

Null Hypothesis Significance Testing is the de facto tool for assessing effectiveness differences between Information Retrieval systems. Researchers use statistical tests to check whether those differences will generalise to online settings or are just due to the samples observed in the laboratory. Much work has been devoted to studying which test is the most reliable when comparing a pair of systems, but most of the IR real-world experiments involve more than two. In the multiple comparisons scenario, testing several systems simultaneously may inflate the errors committed by the tests. In this paper, we use a new approach to assess the reliability of multiple comparison procedures using simulated and real TREC data. Experiments show that Wilcoxon plus the Benjamini-Hochberg correction yields Type I error rates according to the significance level for typical sample sizes while being the best test in terms of statistical power.

When a steering intervention changes a language model's behavior, which part of the intervention deserves the credit? We test this question for a target-centered, scaled conceptor complement by removing its fitted covariance filter while preserving the same target-specific center and gain. At the original aperture α=.05 and gain β=2, all 21 fitted maps across three models and seven targets expand every direction at the generation intervention layers. The matrix correction is 0.093%--7.815% of the matched-control displacement on held-out extraction activations. In 480 paired comparisons within a 1,920-response study, retaining the matrix changes many individual texts but does not establish a consistent additional target-suppression benefit on the four-target panel. A geometric aperture sweep shows that the operating regime changes with the filter setting. Separately, an invertible soft filter preserves what an unrestricted affine readout can recover, whereas dedicated erasure methods reduce prediction by newly fitted probes. Together, the results show why a whole-intervention comparison cannot by itself attribute generated behavior to the fitted matrix: analyze the complete transformation, hold shared components fixed, and test its output consequences directly.

Based on the established task of identifying boosted, hadronically decaying top quarks, we compare a wide range of modern machine learning approaches. Unlike most established methods they rely on low-level input, for instance calorimeter output. While their network architectures are vastly different, their performance is comparatively similar. In general, we find that these new approaches are extremely powerful and great fun.