Papers

Modern vision-language models can accept multi-image contexts, but aggregate accuracy rarely reveals where a multi-image reasoning pipeline fails. We study a controlled schema-application task in which a model must locate two target characters in a 16-image sequence, identify their visual attributes, consult an item-specific visual codebook, and compose two ordered nonce code names. We introduce a within-item 2-by-2 factorial Oracle design: Oracle-A supplies a verified textual schema, Oracle-B supplies a verified ordered target-appearance representation, and both interventions retain the original images and final question. Across 72 new items and 1,188 analyzed model responses, Oracle-B increased strict exact accuracy by 54.6 percentage points (95% confidence interval: 46.3 to 63.0) and outperformed Oracle-A by 60.6 points (Holm-adjusted p-value no greater than .0002). An auxiliary retrieval diagnostic reached only 49.1% ordered-appearance accuracy even after removing code-name generation. Providing both Oracle sources yielded 100% accuracy across all models. These results localize the dominant recoverable deficit to upstream ordered instance recovery rather than primarily to downstream schema use and composition.

Industrial anomaly detection (AD) is bottlenecked by the scarcity of anomalous samples, making high-fidelity anomaly synthesis crucial for improving downstream performance. However, existing methods either rely on scarce real anomaly references or auxiliary data, while lacking complex reasoning and iterative optimization capabilities for generating realistic defects in open-set scenarios. To address these challenges, we propose CORE, a cost-efficient and training-free prompt optimization framework based on Large Multimodal Models (LMMs), which unifies zero-shot and one-shot anomaly synthesis. CORE first introduces an Anomaly Context Grounding module to transform sparse user context into defect-aware generation context, which is subsequently instantiated as structured anomaly prompts that disentangle normal-region preservation from anomaly-specific variation, enabling controllable and faithful defect generation. Furthermore, CORE autonomously refines prompts through a self-supervised contrast-and-reflect mechanism. This design empowers general-purpose LMMs to synthesize realistic anomalies under both zero-shot and one-shot settings without task-specific training. Extensive experiments on MVTec-AD and VisA demonstrate that CORE synthesizes realistic and diverse anomalies, yielding superior improvements in downstream AD performance. Furthermore, its model-agnostic and plug-and-play paradigm enables effortless integration with diverse LMMs, facilitating adaptation to unseen open-set scenarios.

Counts of which brands a large language model names are useful only if they hold long enough to act on. We test this with a pre-registered replication. The same 162 requests (54 questions in English, Estonian and Finnish about three software categories, four of twelve candidate brands per question, three repeats) were sent to gemini-3.1-pro-preview in three waves within about ten hours on 22 and 23 September 2026, and again on 13 October 2026, 20.3 days after the third wave, with identical request bodies. The plan and its decision rule were frozen, hashed and registered on OSF before the first call. The outcome is a lexical mention label per answer and brand, 648 per wave. The primary forecast, each brand-and-question cell's mention rate over the nine earlier answers, scored Brier 0.0281 and log loss 0.126, against Brier 0.0259 for the six-answer forecast of the 8-hour wave and a registered ceiling of 0.0559. It beat brand history (0.104), the overall rate (0.234) and a constant 0.5 (0.250) on both scores and in all nine question families (exact sign-flip p = 0.0039), so durability holds under the registered rule. A forecast frozen from the first wave alone scored 0.0599, 0.0587 and 0.0604 at horizons of 1.5 hours, 9.6 hours and 20.7 days. Of the 202 cells unanimous in the 8-hour wave, 194 were unchanged at D1 and none reversed. The latest wave alone was nearly as accurate as the pooled history (Brier 0.0292; pooled lower in 5 of 9 families, p = 0.28). Every answer reported the requested model version, and none issued a search query although search was enabled. A blind human check of 96 labels agreed with the rule on 94; both disagreements were adjudicated as rater omissions, so after adjudication the rule matched on 96 of 96. The test covers one model, one interval and twelve known brands, and estimates no decay rate.

Rising privacy concerns around large language models have accelerated research in private Transformer inference (PTI). Synthesizing the empirical claims in this line of research, however, is hindered by multi-dimensional cryptographic progress and inconsistent evaluation contracts. In this work, we systematize the field by constructing an empirical graph of 62 PTI systems, mapping system attributes to node features and evaluation compatibility to edges. A three-tiered analysis of this graph illuminates the state of private inference. Node-level attribution shows that cryptographic progress drives performance gains across most systems, yet always co-occurs with other factors (model fidelity, hardware, cost accounting). Edge-level auditing shows that only 1 of 1891 candidate system pairs is directly comparable under shared assumptions. Finally, graph-level analysis shows that the scaling frontier of private inference has plateaued at 13B parameters and projects the cost of deploying current PTI techniques on production-scale Transformers. Overall, this graph-based framework equips the community to evaluate PTI claims transparently and channel technical speedups into deployment-scale private inference.

Long-output inference in large language models suffers from severe KV cache memory and bandwidth bottlenecks. While KV cache quantization alleviates this issue, low-bit quantization (e.g., 2-bit) severely degrades model accuracy. Recent mixed-precision key-cache quantization methods mitigate this degradation by allocating higher precision to important key channels, but estimate channel saliency using only key statistics. Through attention logit perturbation analysis at the RoPE channel-pair level, we show that quantization errors in channel pairs with large query–key pair-norm products contribute significantly to logit variation. Based on this insight, we propose PN-Mix, a mixed-precision key-cache quantization method for RoPE-based decoder-only LLMs, which allocates higher precision to salient channel pairs identified once during the prefill stage and reused throughout decoding. Across standard reasoning, code-generation, and AIME benchmarks, PN-Mix with 2.26-bit average precision achieves comparable or better average accuracy than prior mixed-precision methods using 2.32–2.39 bits. Furthermore, in our Pareto sweep, PN-Mix degrades gracefully down to 1.5 bits, whereas the compared baselines exhibit sharper performance drops below 2.25 bits.

Force feedback provides vision-language-action (VLA) policies with physical observations that can guide action selection and execution. However, differences in tasks, demonstrations, and control settings across existing systems hinder reliable comparisons of force integration designs. We introduce ForceBench, a simulation benchmark for evaluating six-axis wrist-force integration under a unified protocol. ForceBench combines eight tasks spanning physical-property inference, resistance-dependent interaction, and contact-rich assembly with 1,600 synchronized demonstrations and task-specific success criteria. We compare one no-force baseline and six force integration designs on a shared $\pi_{0.5}$ backbone. Prompt Fusion achieves the highest observed mean simulation success of 35.31%, compared with 23.25% for the baseline, and improves success rates across all three task categories. Predictive and temporal gains in simulation concentrate on physical-property inference, while both temporal designs underperform the baseline on assembly. Real-world evaluation of eight systems across five tasks supports the main qualitative trends: Prompt Fusion and Dual-Path Fusion achieve 63.00% mean success, compared with 40.00% for the baseline, but only Prompt Fusion improves every task. These findings support ForceBench's practical value for identifying the benefits and limitations of force integration across manipulation demands. The project page is available at https://forcebench-platform.github.io/.

Transformers have demonstrated strong performance in long-term time series forecasting (LTSF), yet modeling non-stationary temporal dependencies remains challenging. In this paper, we interpret self-attention as a Bayesian posterior over attended positions, combining a content-based prior induced by scaled dot-product similarity with a temporal-evidence likelihood induced by positional encoding. This formulation provides a probabilistic interpretation of how positional encoding shapes the temporal dependencies captured by self-attention. Building on this perspective, we derive a population-level evidence lower bound (ELBO) that characterizes the role of positional encoding in forecasting weakly stationary time series. Our analysis further highlights a limitation of static positional encoding: when temporal dependence structures vary across input windows, a fixed temporal bias must accommodate heterogeneous dynamics and can therefore be suboptimal for non-stationary forecasting. Motivated by this analysis, we propose Bayesian Time-aware Attention (BTA), which employs a mixture-of-experts (MoE) router to adaptively combine temporal kernels and generate input-conditioned, head-specific positional biases. Extensive experiments on standard LTSF benchmarks demonstrate that BTA consistently improves forecasting performance across multiple Transformer backbones and achieves competitive overall performance with modest computational overhead and stable training.

Knowledge distillation (KD) improves language-model pretraining by matching teacher distributions at corpus prefixes, but provides no direct supervision at contexts induced by student predictions. We introduce PA1-OPD, which augments corpus-path distillation with teacher feedback after one student-sampled transition per selected prefix. This expands the supervised states while retaining next-token prediction and corpus-path KD. Data-anchored roots support parallel sampling, and differentiable tree attention shares prefix computation without changing inference. In two continued-pretraining settings, full coverage improves held-out cross-entropy over KD across C4, PG-19, and Pile at every evaluated checkpoint. In Stage2, selecting the 10% of roots with highest student surprisal adds 48.85% to KD's Macro cross-entropy gain over next-token training. Gains remain when branch supervision excludes successor states already on the corpus path, supporting the value of feedback beyond observed contexts. Additional depth or breadth improves aggregate results at unchanged per-term supervision weight, with benefits varying by corpus. These findings support local student-induced supervision as a complement to pretraining distillation.

Recent vision-language-action (VLA) models have shown strong performance on local robot manipulation, yet their success does not readily translate to long-horizon mobile manipulation. In our experiments, a pretrained subtask VLA achieves 80% success rate on atomic manipulation tasks but only 19% when these skills are composed into mobile manipulation. We argue that this gap stems from failures across the execution process rather than from local manipulation capability alone. To uncover these bottlenecks, we systematically diagnose failures in long-horizon mobile manipulation and identify three dominant regimes: task planning, navigation and base positioning, and local manipulation. Based on this diagnosis, we develop failure-guided agentic intervention that augments a fixed subtask VLA with demonstration-based planning guidance, adaptive visual memory, accessibility-aware base adjustment, and local end-effector adjustment. On RoboCasa365, these interventions progressively improve task success from 19% to 47%, with the finalized system reaching 49% on held-out episode seeds, compared with 18% for the baseline. The same design improves performance from 14% to 37% on BEHAVIOR-1K and from 5/20 to 13/20 on real-robot tasks. These results suggest that reliable mobile manipulation can be advanced by diagnosing and intervening on the current bottleneck, rather than uniformly scaling the underlying policy. Our code will be released at https://github.com/PKU-VaLuE-Lab/Cracking-Mobile-Manipulation.git.

A coding agent repeatedly makes decisions as it repairs code. These decisions shape repair trajectories and affect the token cost of successive model requests. We ask which repair decisions can be delegated to a low-cost probabilistic judge to reduce token use while maintaining repair quality. The judges answer questions with probabilities over predefined answers, without generating explanations. We evaluate judgment quality across repair stages and test how individual interventions change subsequent repair work. We compare these interventions with unguided repair, which uses no judge, and with the same workflow using the repair model itself as the judge. The benefit depends on the work an intervention changes, rather than the cost of the judgment alone. Based on these studies, we propose JevCoder, a workflow built around Jev, a low-cost probabilistic judge. It combines code-snippet recommendations with action-type guidance and uses judgments to determine when to submit. We evaluate three repair models, each on the same 400 tasks: 200 from SWE-bench Verified and 200 from SWE-Bench Pro Verified. Compared to unguided repair, JevCoder reduces mean repair-model token use by 21.6%-50.6%, with higher observed repair rates across all three models.

Whole-body mobile manipulation suffers from two fundamental challenges: inherent motion heterogeneity across locomotion, arm, and hand modules, and progressive error accumulation over long-horizon multi-stage tasks. Existing teleoperation platforms also lack integrated dexterous tactile sensing to support coordinated whole-body decision-making. This paper presents RoboSpan, a high-DoF whole-body teleoperation and demonstration platform, and the Hierarchical Modular Framework (HMF) for learning modular coordination policies. HMF functionally decouples the 29-DoF action space into fine hand manipulation, arm pose adjustment, and mobile-base spatial configuration branches, and introduces the Coordinated Modular Embodied Transformer (COMET) with inter-modular cross-attention to enable dynamic information exchange and action coordination. Vision-language model driven hierarchical task decomposition structurally suppresses temporal error propagation, while a full-palm 2D tactile representation encodes distributed contact signals for cross-module tactile empowerment. Experiments on a physical 29-DoF robot demonstrate that the proposed method achieves over 20% higher task success rate than mainstream baselines across a range of long-horizon whole-body tasks, markedly improving stability and robustness in contact-rich manipulation.

The ability to generate proteins with desired functions is a central goal of computational biology. Although natural language can express diverse design objectives, existing protein generation methods often support only narrow tasks or make limited use of biological constraints. We develop InstructPro2 to translate natural-language design objectives into protein sequences while incorporating task-specific biological context. A text encoder captures the design instructions, while dedicated ligand and antigen encoders provide biological context to the sequence-generating decoder. For pretraining, we assemble InstructPro2Bench, comprising 205.7 million <function description, protein> pairs from UniProtKB. We subsequently adapt the model to ligand-binding protein and antibody design. On functional protein generation benchmarks, InstructPro2 outperforms Pinal, the strongest baseline, achieving a functional recovery F1 of 74.8 versus 62.2 and lowering global RMSD from 5.33A to 3.04A. For antibody design, the model obtains a DockQ score of 0.39, while ligand-binding protein design reaches a docking success rate of 72.6%. These findings highlight how integrating language-based instructions with biological conditioning can extend protein generation to distinct design settings.

The lasting gains in AI have come from methods that absorb more computation, and language models already scale computation within a call. Useful work, though, happens across calls: over hours or days, through tools and external systems, among many agents and the humans they work for, and increasingly through code the agents write themselves. Whether that work compounds is enabled by the harness around it, and the harness faces a constraint the model does not: of the three participants in such a computation, the human, the model and code, only the human's attention does not scale, so the computation can grow only if the human is not in every loop. L3M is a runtime built on that premise, written in Lean 4 so that one language implements the harness, states its invariants, and is the language agents extend it in. One computation spans humans, agents, models and code, and all of its state lives in Git: a rich state trajectory – rather than a chat conversation – is an object the runtime can resume, rewind, fork, compare and merge. Waiting, communication, delegation and undo are primitives, not patterns the model reasons through each time. When an agent works out a procedure, it can write it once as a sealed Lean program whose type is its permission request, and the harness offers it back as a tool – machinery for the model to use. To take the human out of the loop, L3M uses reversibility to decide where that is safe: a tool's review tier follows from how far its effects can be undone, so agents act freely where the runtime can recover. Because the machinery can change, the properties that make this safe are stated as kernel-checked theorems about the code that runs. On Terminal-Bench, an untuned L3M reaches about two thirds of its (unsafe) shell-equipped pass rate using only safe (not needing human review) tools.

Repeated event forecasting requires extracting predictive signals from noisy evidence and evaluating their value across successive decisions. We introduce JockeyBench, a benchmark and evaluation protocol that studies this problem through historical Japanese horse racing. The protocol combines time-gated evidence, priced tools, paired frozen-program controls, and two workflows: executing an LLM-written predictor or acquiring evidence for a fresh per-race LLM forecast. Evaluation separates forecast quality, benchmark-specific net utility, and retrospective betting returns and risk under fixed decision rules. Across 28 model–scaffold configurations on 324 races, paired tests identify seven revision gains and two declines after multiple-testing correction. Yet every configuration has negative mean profit under capped-Kelly settlement. In a four-model comparison, the reusable-program workflow has higher mean betting profit and lower research fees, although both workflows lose on average. Further diagnostics reveal that mean probability bias reverses on selected bets and that profitable individual runs can depend on concentrated returns. These findings motivate evaluating predictive competence, adaptation, resource use, and decision value separately when assessing agents for repeated event forecasting.

We consider the problem of estimating the background intensity of a Hawkes process when the triggering mechanism is known, treating the background as an unknown finite positive measure with a regular density. We propose a likelihood-based variational approach based on the Wasserstein-Fisher-Rao geometry. In this setting, the relevant first variations are naturally distribution-valued and therefore fall outside the classical gradient-flow formulation. We develop a distributional framework for the resulting dynamics and give them a finite-dimensional meaning through projections onto statistical manifolds. We identify the true background as the unique zero of a limiting generalized gradient and, under suitable stability, compactness, integrability and approximation conditions, establish a conditional consistency theorem for measurable estimators satisfying the empirical projected equations on events whose probabilities tend to one. Finally, we specialize the construction to Gaussian-mixture manifolds and obtain an explicit finite-dimensional scheme.

Social psychology has long described how people can be persuaded to interpret a harmful act as moral. Recent research suggests that Large Language Models (LLMs) echo aspects of human psychology, from cognitive biases to social conformity. This raises a question with direct safety implications: can LLMs be moralized to bypass safety guardrails? We answer this question with a multi-agent framework grounded in Moral Foundations Theory (MFT). Given a harmful request that is initially refused, our agents diagnose which moral foundation (MF) the prompt is violating, identify a justifiable foundation under which the same request appears virtuous, rewrite the prompt accordingly, and reinforce it with generated precedents of comparable moral reframing. We evaluate our framework on harmful prompts from three well-known jailbreak datasets across five open-source and proprietary models. Our analysis reveals that models are highly susceptible to moralization, exhibiting high single-trial jailbreak rates with success increasing further under repeated attempts. We also find that safety benchmarks are imbalanced in the MFs they violate, concentrating on foundations toward which models already exhibit biases.

We construct a seven-cycle example for a periodically forced scalar equation with a continuous three-zone piecewise-linear nonlinearity. Symbolic integration and linear elimination reduce the four-flight matching conditions to two equations in the residence times. The competing exterior visits produce an additional turn of the critical response and lead to the rational parameters $(A,B,\mu)=(9,3/2,181363/100000)$. At these parameters, finite Taylor estimates give three sign changes of the return displacement. Scalar order, analyticity along transverse itineraries, and half-period reflection then establish at least seven distinct isolated periodic orbits on the phase cylinder.

Greibach and Ginsburg (1969) studied abstract families of languages or AFLs. A full AFL is a class of languages closed under the following operations: homomorphism, inverse homomorphism, intersection with regular languages, union, concatenation, and Kleene star. Greibach (1970) used a restricted substitution, called nested iterated substitution, to introduce the notion of super AFLs, which are full AFLs that are further closed under nested iterated substitution. The smallest super AFL was shown to be the class of all context-free languages. From this fact, we focus on Greibach's limited, univariable application of nested iterated substitution and further generalize it into k-variate. This introduces the notion of nested iterated $k$-variate substitution whose definition does not rely on "machines". Recently, a "two-tape" version of Hibbard's (1967) rewriting systems, which were later reformulated by Pighizzini and Pisoni (2014) as k-limited automata with "single" input/work tapes, has been studied under the name of depth-k storage automata. This work unearths a few characteristics of the languages recognized by those automata under a restriction known as frozen blank sensitivity in terms of full AFLs and nested iterated k-variate substitution. This will promote a further abstract treatment of those languages.

As large language models (LLMs) are increasingly used to generate text, their misuse has become a growing societal problem, with implications for academic misconduct, model distillation, and cyberattacks. For these reasons, Model Authorship Attribution (MAA), which identifies the source model of a given text, is important for promoting the responsible use of LLMs and improving AI safety. In practice, attribution systems must remain reliable beyond the distributions observed during training. They may encounter texts from unseen domains and writing settings, as well as generations from future versions of model families that were not available when the system was trained. However, existing MAA methods often exhibit substantial performance degradation under such distribution shifts. To address this challenge, we propose Authorship Tracing via Likelihood-based Attribution of Sources (ATLAS), which formulates MAA through source-conditioned text modeling rather than direct source prediction. ATLAS learns how text from each source family is generated and attributes a given text by comparing its likelihood under each learned source condition. We aggregate this likelihood evidence at the token level with clipping so that no single token dominates the decision. We further introduce MAABench, a broad MAA benchmark for evaluating practical OOD settings such as domain and version. Our experiments show that ATLAS consistently outperforms existing baselines across MAABench and OpenTuringBench. In particular, when both the generation domain and the model version change on MAABench, ATLAS achieves 74.27% accuracy, 9.00pp higher than the strongest baseline. Beyond MAA, ATLAS can also help identify distillation source models by measuring which source model a target LLM most closely resembles. Our extended experiments show that ATLAS recovers the known teacher of every distilled model we evaluate.

Large language model (LLM) agents can reuse experience from prior task executions to improve performance on subsequent tasks. Although existing methods enable cross-agent experience sharing, differences in agents' execution histories and task requirements make shared experiences difficult to reuse effectively. We propose XferAgent, a framework for cross-agent experience transfer that organizes reusable experience in local and shared experience graphs. XferAgent reconciles compatible contributions from heterogeneous agents while avoiding unreliable merges. For each new task, it selects shared experiences and integrates them with the target agent's local experience to support task execution. Experiments across diverse task domains show improvements over representative memory-based methods and effective experience transfer between heterogeneous agent frameworks. Cumulative task success also increases over successive task passes as agents update and reuse their experiences.

State-of-the-art visuomotor policies are increasingly capable of executing complex tasks, yet can fail during deployment in ways they cannot predict or recover from on their own. Improving their performance and robustness typically requires collecting more data, retraining, or modifying the policy architecture, all of which are time- and resource-intensive in practice. We instead propose an offline-trained monitoring-and-intervention framework that improves a policy during deployment using only black-box query access. From a limited set of labeled on-policy rollouts, we learn a policy-conditioned liveness value that measures the nominal policy's progress toward task completion; we additionally train a compact latent world model using these rollouts together with play data. At deployment, the liveness value continuously monitors execution. When it detects a loss of progress, a short-horizon planner searches the world model for corrective actions using the same value as its planning objective, after which control returns to the nominal policy. We evaluate our method on the LIBERO-Goal task suite and on a Franka Panda manipulator, using the vision-language-action model $\pi_{0.5}$ and a diffusion policy as nominal policies, and demonstrate significant improvements in task success rates, while supporting real-time control. Additional results and videos are available on the project website (https://policy-improvement-wm.pages.dev/).

Graph Collaborative Filtering models have achieved remarkable success in recommendation systems. However, they inherently suffer from a catastrophic representation collapse in strict cold-start scenarios due to the topological isolation of unobserved items. While integrating multi-modal semantic features offers a potential remedy, directly mapping isolated semantic attributes into a globally topology-driven collaborative space frequently precipitates representation degradation and optimization conflicts. To bridge this fundamental structural divide, we propose DualGCN, an innovative Dual-Branch Graph Knowledge Distillation Framework. Operating under a decoupled two-stage training paradigm, DualGCN first constructs a robust collaborative teacher network where item representations are anchored by a topology-smoothed semantic prior, effectively reshaping the latent manifold. Subsequently, a semantic student branch employs a residual mapping mechanism and a magnitude-agnostic cosine distillation objective to achieve high-fidelity cross-modal alignment without disrupting established collaborative signals. Furthermore, we introduce a dynamic norm reconstruction mechanism to rigorously calibrate the physical scales of novel and warm items during global retrieval. Extensive experiments across three real-world benchmark datasets demonstrate that DualGCN significantly outperforms state-of-the-art baselines in strict cold-start scenarios while comprehensively preserving recommendation accuracy for highly active items.

Temporal graph data are recorded in different representations, but whether these differences also determine the scope of learned knowledge reuse remains unclear. This study investigates whether predictive structure learned from relational histories can be reused between continuous-time dynamic graphs (CTDGs) and temporal knowledge graphs (TKGs). Bidirectional reuse emerges while retaining distinct input and prediction interfaces and independent vocabularies, and sources with the same representation are not always more useful. Controlled reorganizations of source histories show that the organization contributing to reuse varies across source-target conditions. We show that learned aspects of Relational Dynamics, the predictive structure present in relational histories, can be reused across conventional representation boundaries. Independently pretrained TKG and CTDG learners can also be fused at the score level to improve prediction on a third dataset without entity or relation-semantic alignment. Two proofs of concept in recommendation and security detection illustrate that external temporal sources without direct alignment can also become useful pretraining resources.

Differential Privacy (DP) provides a rigorous framework for quantifying privacy guarantees. While most methods achieve DP by injecting calibrated random noise, the intrinsic randomness of certain procedures such as sampling can yield DP guarantees “for free”. In this work, we develop a unified Rényi divergence framework to characterize the inherent DP guarantees of sampling from a Bayesian posterior distribution. We show that the privacy loss incurred by One-Posterior-Sampling (OPS) is governed by posterior exponential moments of the single-record log-likelihood ratio, and establish explicit DP guarantees under uniformly bounded, sub-Gaussian, or sub-exponential tail regimes. We apply the framework to representative Bayesian models, including categorical likelihoods with arbitrary priors, Gaussian-Gaussian conjugate pair, generalized linear models from the exponential family and linear regression with Gaussian likelihood, both with Gaussian priors on the regression coefficients. Our results elucidate sample size, model structure, neighboring relations in DP (bounded or unbounded), and how prior if applicable, jointly determine the inherent DP guarantees of OPS. Furthermore, our results recover the existing DP guarantees for OPS as special cases – often with tighter privacy loss bounds – and substantially broaden the class of Bayesian models for which the inherent DP guarantees of OPS can be rigorously characterized.

On-policy distillation (OPD) transfers knowledge from a teacher model to a student along student-generated trajectories, typically by minimizing the reverse Kullback-Leibler (RKL) divergence. However, RKL can provide weak corrective signals for teacher-supported tokens to which the student assigns very low probability. This raises two questions: which positions are most consequential for the behavioral gap between the student and teacher, and what additional supervision is needed at these positions? Through cross-sampling, we find that divergence-based routing more effectively pinpoints positions contributing to the teacher-student performance gap than entropy-based routing. Further analysis reveals teacher-mode under-coverage at high Jensen-Shannon (JS) divergence positions, where standard OPD increases the large teacher-student probability gaps. Motivated by these findings, we propose Divergence-Routed On-Policy Distillation (DrOPD). DrOPD retains RKL supervision at all positions and selectively adds forward Kullback-Leibler (FKL) divergence supervision at positions with high JS divergence. Experiments across six math benchmarks show that DrOPD improves over standard OPD and entropy-based routing. DrOPD also reduces severe under-coverage cases by 37.9 percent relative to OPD, supporting targeted FKL as an effective complement to OPD.

Real-world engineering drawings are often large-format and information-dense. Answering practical engineering questions therefore requires both fine-grained understanding of engineering information and complex reasoning over objects, relations, and evidence distributed across different regions or views. Existing benchmarks cover a range of engineering-drawing understanding tasks involving local regions, individual views, specific elements, or more constrained visual settings. However, systematic evaluation of understanding over realistic, large-format engineering drawings remains limited. To address this gap, we introduce LED-BENCH, a multidisciplinary benchmark designed to evaluate fine-grained understanding and complex reasoning over large-format engineering drawings, spanning the mechanical, architectural, and transportation domains. Extensive evaluation of 11 representative multimodal large language models, including 8 proprietary models and 3 open-source models, reveals substantial remaining difficulty: GPT-5.6 Sol achieves 52.65% average accuracy, performance varies across engineering domains, and all evaluated models perform worse on questions requiring broader evidence integration. These results highlight fine-grained understanding, evidence integration, and complex reasoning over large-format engineering drawings as persistent challenges for current multimodal models.

Quantifying the diversity of a set of objects is a fundamental problem in machine learning and statistics. Mironov and Prokhorenkova (2025) recently showed that three axioms-monotonicity, uniqueness, and continuity-are crucial for a reliable diversity measure and that none of the standard metrics in the literature satisfy all three. They also constructed two measures that do satisfy the axioms, but both are NP-hard to compute, leaving open the question of whether efficient axiomatic diversity measures can exist. We study a Euclidean-restricted version of this question. We introduce Geometric Kernel Diversity (GKD), which associates a radial kernel solid with each point and measures the volume of their union. We show that canonical GKD satisfies all three Mironov-Prokhorenkova axioms on Euclidean-realizable inputs. For canonical GKD, we give exact polynomial-time algorithms in dimensions $d \le 3$. In arbitrary dimensions, we give a randomized algorithm that returns a $(1 \pm \varepsilon)$-approximation with probability at least $1-\delta$ in time $O \left(\frac{nd}{\varepsilon^2}\log\frac{2}{\delta}\right)$ for $n$ points. The axiomatic guarantee applies to exact canonical GKD; the randomized algorithm approximates this deterministic quantity. We further show that the two diversity measures of Mironov and Prokhorenkova (2025) remain NP-hard on Euclidean-realizable inputs and, unless $\mathrm{NP} = \mathrm{RP}$, admit no FPRAS even on this restricted domain.

Latent reasoning replaces natural-language chains of thought with compact continuous "soft-token" representations, reducing the number of intermediate tokens that must be decoded autoregressively. We use a small external encoder to induce latent reasoning in a frozen backbone model. The encoder iteratively maps the backbone's hidden states to new latent blocks, constructing a latent reasoning prefix, after which the frozen model writes its remaining reasoning and its answer as text. Our method, INTACT, trains this encoder through distillation to reproduce the effect of written reasoning steps while keeping the model's behavior close to that of the frozen backbone, and then reinforces it on answer correctness. Across four frozen backbones from 1B to 8B, INTACT improves both GSM8K and average out-of-domain accuracy, including GSM8K gains from 75.8% to 79.3% on Llama-3.2-3B and from 77.6% to 79.7% on Qwen2.5-3B. It also reduces end-to-end latency per problem by 19% on Llama-3.2-3B and outperforms the evaluated latent-reasoning baselines at the 3B and 8B scales. Beyond the mathematics setting, the same recipe trained on a multi-domain corpus improves the frozen backbone on twelve evaluation benchmarks spanning mathematics, science, and commonsense, without task-specific adaptation.

Autonomous drones are increasingly deployed in safety-critical tasks, yet their autopilot software can still contain subtle logic flaws that lead to unsafe behaviors. These flaws are difficult to discover with existing techniques because they often depend on semantic relationships across modules, commands, and runtime state, and rarely manifest as crashes or detectable failures. With recent advances in LLMs, we observe that, when equipped with essential domain knowledge and carefully crafted context, LLMs demonstrate strong capabilities in understanding the high-level semantics of autopilot systems, making them suitable for logic flaw discovery. However, LLMs are limited by their context and can be biased to the vulnerabilities frequently occurs in training dataset, while logic flaws often require higher-level semantic bug oracles to identify their existence. To bridge the gap, we present Impossible Mission Finder (IMF), a multi-agent system combining static analysis, dynamic execution and log analysis, that supports fully automated identify, test, and validate logic flaws in autonomous vehicles. IMF primarily contains three LLM agents, it first uses the auditor agent to propose logic flaw hypothesis based on carefully-crafted program slicing, then it launches tester agents to write and execute test cases for each candidate, and finally it spawns the log analyzer agent to validate the execution log against the flaw hypothesis. This enables IMF dynamically validate bug oracles in a case-by-case manner, making it suitable for discovering logic flaws that do not have a unified pattern. We evaluated IMF on two widely used open-source autopilot systems, PX4 and Ardupilot, leading to 105 previously unknown logic flaws, that can cause unsafe behaviors including mission interruption failures, invalid actions, entering into restricted fly zones etc.

The rapid growth of AI data centers introduces large, workload-dependent loads whose protection and reconnection behavior can affect power-system transient stability. Calibrating dynamic models of such loads is challenging because treating all associated elements as freely estimable may lead to weakly identified or physically confounded parameterizations. This paper proposes an identifiability-guided calibration framework for a composite load model with separate IT, cooling, and auxiliary branches. A nuisance-projected Fisher-information criterion characterizes the local identifiability of the cooling-motor composition, while a workload non-equivalence result shows that voltage-only recalibration cannot reproduce workload-dependent responses. Guided by these results, a grey-box estimator trained with branch-level supervision learns the workload-to-power relation and the cooling composition, while retaining a prescribed UPS protection law and an externally anchored reconnect-ramp duration. The learned workload relation is then distilled into a closed form for transient-stability simulation. Evaluation using 300,000 MIT Supercloud GPU power-utilization samples and 12,600 IEEE 39-bus fault scenarios yields an aggregate active-power $R^2$ of 0.948, compared with 0.856 for an LSTM baseline. Removing branch-level supervision preserves relatively high aggregate accuracy but substantially degrades cooling-component recovery. Under unseen deep voltage sags, the proposed model achieves a transient $R^2$ of 0.750, compared with 0.033 for the LSTM. Closed-loop simulations further show that cooling composition and protection behavior can affect retained-load states and voltage responses near protection thresholds.