Papers

A transformer can make an attribute linearly decodable in its residual stream at a depth where that attribute does not yet influence the output. This gap between where information is readable and where it is used has been shown for attributes stated directly in the input. We ask whether it also holds for an attribute the model must infer gradually over a conversation, namely how expert its dialogue partner is. Using ExpertCollab, a corpus of multi-turn research-planning dialogues between model-played personas at four expertise levels, we find that partner expertise is most decodable in the early layers and falls to near chance before the midpoint of the network. Counterfactual patching shows that injecting the expertise difference at the layer of peak decodability barely changes a fixed late-layer readout, whereas the same difference injected past the midpoint propagates almost completely, a separation of more than an order of magnitude. A content-matched random control and a probe-free diagnostic place the transition at the same early layer, and a statically specified control attribute stays decodable throughout. An inferred relational attribute is therefore represented well before it becomes causally active, which bounds where any attempt to read out or steer partner-conditioned behavior must intervene. We use one model on a synthetic corpus as an initial demonstration.

Recent image-to-3D generation models built on flow-matching diffusion Transformers (DiT) can produce high-fidelity meshes, yet their post-training strategy remains largely unexplored. There exist several critical bottlenecks in reinforcement learning: the inherent difficulty of defining comprehensive rewards for 3D geometric quality, and the gradient interference that arises when jointly optimizing heterogeneous objectives. Inspired by the practicability of on-policy distillation (OPD) in large language models and image generation, we propose Flow3D-OPD, a two-stage post-training framework that introduces multi-teacher distillation into 3D geometry generation. In the first stage, we utilize the semi-policy to enhance the foundational capability of the pretrained model and then design an agentic verifier for 3D geometric quality evaluation. Based on the verifier, we could cultivate domain-specialized teacher models via direct preference optimization (DPO). In the second stage, we consolidate heterogeneous expertise into a unified student model through on-policy distillation with hard task-routing sampling and gradient accumulation, which could mitigate the gradient interference in joint optimization. Without relying on elaborate modifications, our straightforward yet effective design achieves consistent improvements across all geometric quality dimensions and surpasses all teacher models in the average metric. Extensive experiments demonstrate that our approach provides an effective paradigm for reinforcement learning in 3D generation.

Large language model (LLM) agents increasingly achieve long-horizon tasks by combining foundation models with explicit skills and implicit procedural knowledge acquired through execution. The resulting task-solving capabilities have become valuable proprietary assets, raising a new security question: can a substantially weaker attacker-controlled agent acquire the capabilities of a stronger proprietary agent through limited black-box interaction? Existing skill-stealing attacks recover explicit skill artifacts, yet we show that artifact leakage does not necessarily transfer capability: a weaker agent may possess the same skills but still fail because it lacks procedural behaviors implicitly realized by the stronger agent. Our key insight is that the skill execution gap itself forms a leakage surface, where missing behaviors are exposed through observable differences between successful victim executions and failed attacker executions. Based on this, we present AgentLeak, a black-box capability-cloning attack that identifies capability-critical behaviors from these execution differences and incorporates them into attacker-side skills, while keeping the attacker's model, harness, and tools unchanged. Across 20 task scenarios comprising 600 instances, diverse agent systems, and multiple backbone models, AgentLeak improves task pass rates by over 40% compared with direct skill reuse and recovers more than 80% of the victim–attacker capability gap. Our findings reveal a confidentiality risk in LLM agents: protecting explicit artifacts alone is insufficient, as observable execution behavior can leak the procedural knowledge required to reconstruct proprietary task-solving capabilities in low-capability and attacker-controlled agents.

Reliable risk assessment remains a central challenge for Autonomous Vehicles (AVs). Despite advances in automation, passenger cognition provides a non-intrusive auxiliary signal that improves both objective and perceived safety without requiring active human intervention. We introduce an Electroencephalogram (EEG)-based Brain-Computer Interface (BCI) that decodes passenger neural responses for both Risk Prediction (RP) and Danger Identification (DI), explicitly modeling humans as passengers to match real-world AV use. To achieve this, we propose the Passenger Cognitive Model (PCM), Risk-aware Sequential Labeling (RSL), and the Passenger EEG Decoding Strategy (PEDS), which integrates a 3D Convolutional Recurrent Neural Network (3D-CRNN) model for joint EEG decoding. Experimental results show that 3D-CRNN achieves a Balanced Accuracy (BA) of $95.3\% \pm 2.7\%$ in RP and improves single-subject DI from $80.9\% \pm 3.9\%$ to $85.0\% \pm 3.2\%$ with RSL. Event-wise analyses further show that 3D-CRNN consistently outperforms other models across different event types in RP and DI. In generalization experiments, 3D-CRNN achieves $77.0\% \pm 5.3\%$ BA in cross-session DI and $77.4\% \pm 1.1\%$ BA on seen subjects in cross-subject evaluation, while maintaining a $64.9\% \pm 8.5\%$ BA on unseen subjects, demonstrating promising generalizability and transferability across both intra-subject and inter-subject variability. These findings establish an Electroencephalogram (EEG) decoding framework for AV passenger hazard perception and suggest that passenger cognitive signals can provide auxiliary supervision for future AV decision-making and Safety of the Intended Functionality (SOTIF) support.

Quadruped robot locomotion policies are often trained using reinforcement learning, which in turn relies heavily on hand-crafted reward functions. Designing reward functions requires substantial manual engineering, and it is often unclear which local rewards will induce the desired global behavior. Shaped rewards from formal specifications in languages like Signal Temporal Logic (STL) can make rewards more interpretable, but writing STL specifications itself still requires domain expertise. We study whether large language models (LLMs) can fill this gap by generating Parametric Signal Temporal Logic (PSTL) specifications that are subsequently used for policy learning. Given a natural language locomotion objective and a constrained specification grammar, GPT-5.5 and Qwen 3.6 independently propose STL templates for command tracking, safety, and gait structure. We instantiate the parameters of the generated PSTL templates using expert trajectories and retain only specifications that are consistent with demonstrated expert behavior. The resulting specifications are then transformed into smooth, finite-history reward functions and used to train a quadruped locomotion policy with Proximal Policy Optimization (PPO) in MuJoCo XLA (MJX). We evaluate both gait-aware and gait-agnostic settings. The former specifies walking-trot, trot, and bound regimes, while the latter allows contact patterns to emerge from the task objective. We compare against hand-engineered rewards, Text2Reward-style LLM-generated reward code, and an expert-switching oracle. Gait-aware Qwen 3.6 specifications achieved 100% survival and command success across all tested speeds (0.3–2.1 m/s) and matched the target gait at high speeds, whereas Text2Reward achieved 0% for both metrics at $\geq 1.9$ m/s. Videos: https://stl-locomotion.github.io/

Natural adversarial examples (NAEs) reveal that vision models can fail under realistic semantic changes beyond norm-bounded perturbations. However, generating NAEs in a black-box setting remains challenging because existing generative attacks often rely on surrogate models, learned attack priors, or costly query-based optimization, whereas the natural transformations that expose model vulnerabilities are unknown a priori. We propose Adversarial Scenario Attack (ASA), a query-based black-box framework that searches over natural-language editing scenarios using a multimodal language model and a modern text-guided generative editor. ASA jointly explores background, weather, and material/color transformations through winner–loser feedback, and uses a greedy explorer to compose only attack-improving scenarios. Across diverse ImageNet classifiers, ASA achieves substantially higher attack success rates than prior query-based generative attacks while requiring fewer victim-model queries and preserving competitive perceptual quality. Moreover, ASA exhibits both image-level and prompt-level transferability: its adversarial images remain effective across victim-model architectures, while its discovered editing scenarios can be reused across same-class images and, in some cases, across architectures. These findings suggest that vision models possess reusable vulnerabilities to natural transformation patterns, which ASA can efficiently identify in a black-box setting.

Micro-dramas, characterized by ultra-short durations and hyper-dense storylines, pose unique challenges for video understanding that conventional benchmarks fail to address. To bridge this gap, we introduce M-Drama, the first large-scale bilingual benchmark for micro-drama comprehension, featuring over 35K instances across 9,138 clips. Furthermore, while reinforcement learning can enhance VLMs on complex narratives, existing reward metrics often suffer from sparse and superficial signals, failing to capture intricate character identities and temporal structures. We propose SAGA (Structure-Aware Graph Alignment), a novel graph-matching reward function that models narratives as heterogeneous graphs. SAGA computes dense, rigorous rewards via decoupled semantic triplet and structural temporal matching. Extensive experiments on Qwen3-VL-8B-Instruct demonstrate that SAGA outperforms existing baselines, delivering substantial improvements in open-ended accuracy and summary quality, while maintaining competitive out-of-domain generalization. Code is available at https://github.com/qyx1121/MDrama_SAGA.

Astronomical spectra encode rich physical information, but drawing scientific conclusions from spectral features typically requires expert interpretation. This paper presents AstroSpecLM, a spectrum-language model that connects one-dimensional DESI spectra with Qwen3-4B to answer questions and provide explanations grounded in spectral evidence. Instead of generating question-answer pairs directly from templates or raw catalog fields, we first distill each spectrum into a compact set of catalog- and spectrum-derived facts, then use these facts as references to generate instruction-following conversations. The resulting model is competitive with specialist supervised baselines on classification and redshift estimation, while additionally producing natural-language explanations that reference specific spectral features. Our results indicate that grounding a language model in one-dimensional scientific spectra is feasible, and that fact-mediated instruction data yields a model capable of both prediction and explanation.

Credit card fraud detection typically relies on tabular features, while repeated attributes can also provide useful relational signals. This paper proposes THGT-FD, a Temporal Heterogeneous Graph Transformer for Fraud Detection. Each transaction is represented using one transaction token and six types of relation tokens and incorporates Time2Vec encoding into the transaction representation. A Transformer learns the interactions among these tokens within each individual transaction and then outputs a fraud probability. Experiments were conducted on 150,000 transactions sampled from the IEEE-CIS Fraud Detection dataset and chronologically partitioned according to TransactionDT. On the test set, THGT-FD achieved an AUC-ROC of 0.8536, an average precision of 0.4164, and a Recall@5% of 0.4708. The class-weighted histogram-based gradient-boosting baseline achieved an AUC-ROC of 0.8722. The results indicate that relation tokens provide useful information for fraud-risk ranking, although the current model does not yet incorporate entity-level historical aggregation.

We present our submission to the EgoProactive track of the ECCV 2026 Wearable AI Challenge, which ranked first in the large-model division and second in the <=2B division. The task requires a wearable assistant to decide after each eight-second segment of egocentric video whether to intervene or remain silent. Our approach has two main components. First, we reformulate intervention timing as single-token classification. Rather than generating either $interrupt$<utterance> or $silent$, the model predicts yes or no, and we derive the decision from the renormalised probabilities of these two tokens. This formulation improved macro-F1 by 0.249 and G-mean by 0.30 over free-form generation. Second, because labelled data were limited to the released validation set, we generated additional supervision using a tool-calling video agent that inspects each clip and assigns intervention timestamps. A narration-only alternative was four times larger and ten times cheaper, but transferred worse than supervision from an unrelated real corpus, suggesting that visual grounding is more important than annotation volume for this task.

LLM assistants often produce more answers than humans can review before users see them. Most evaluations ask whether an answer is wrong, unsupported, or low-confidence. Bounded review budgets instead ask which answers should be checked first under a fixed review budget. Risk alone is not enough: a high-risk answer may be hard to repair, while a moderately risky answer may be directly correctable from available evidence. For generated-answer evaluation, we model review prioritization as exposure reduction, where review value combines estimated wrongness, intervention affordance, impact, and cost. We evaluate review queues with Wrong-Answer Exposure Ratio (WAER), the fraction of wrong answers left unreviewed, and post-repair residual exposure (PRRE), the fraction still exposed after deterministic benchmark-supported repairs. PRRE uses repairability rules that do not numerically reuse the affordance scores used for ranking. On a 720-item TAT-QA/SciFact stress benchmark, review-value ranking keeps answer-level WAER nearly unchanged at 20% budget (0.605 vs. 0.600) but lowers PRRE from 0.881 to 0.716. These results show that trustworthy LLM evaluation should measure not only error detection, but also how limited review capacity reduces exposed wrong answers.

Retrieval-augmented generation (RAG) is commonly evaluated by whether the final answer is correct. That test is insufficient: an answer can match its reference while the context that produced it contains a direct contradiction, leaving the contested evidence invisible to answer-only review and retrieval relevance scores. This paper presents the Hierarchical Consistency Framework (HCF), a post-hoc, model-agnostic audit of three distinct levels of a RAG process: the knowledge corpus, the final retrieved context, and the generated answer. HCF represents corpus conflicts as source-linked atomic facts, thereby identifying the documents responsible, and returns each Answer Consistency Score (ACS) with an explanation of supporting and contradictory contextual statements. We evaluate HCF on several controlled corpora spanning five domains and 100 query-corpus instances. A human evaluator compares every generated response with its supplied ground-truth response. The results show that the three diagnostic levels can dissociate: the corpus with the highest mean retrieval similarity has the lowest mean ACS, while a structurally degraded corpus performs worse at corpus level but better at answer level. Most importantly, HCF identifies contradictory retrieved evidence in several cases where the answer still matches the ground truth. HCF does not certify factual truth; it makes the evidence supporting and challenging an answer inspectable and attributable.

Agent-to-agent (A2A) alpha discovery is slowed by repeated feedback cycles between mining and evaluation agents, whose hand-offs, in contemporary LLM multi-agent systems, are free-form natural-language messages that carry no stable contract and cannot be replayed. We first restructure this communication as a structured agent-to-agent protocol of typed, causally addressable, unicast records, so that the committed stream forms a causal trajectory. On that trajectory a single predictor with four typed heads forecasts the accumulated guidance the two miners would receive several cycles ahead; a transactional verify–leap controller then commits a multi-cycle speculative outcome only when it passes a four-level gate, and otherwise rolls back to the exact prior state. Structure is the enabling contribution, and its value is not accuracy. A controlled ablation shows an equal-information free-text channel reaches the same predictor hit rate. What typing provides is a state that can be schema-checked, replayed deterministically, and prevented by construction from leaking a forecast to an evaluator: auditability by construction, not an empirically stress-tested guarantee. On a CSI~1000 out-of-sample holdout, our single run is the only one among eight methods (seven baselines and ours) to hold a positive median annualized return and Sharpe at the factor level, though the median return in excess of the benchmark stays negative for every method including ours; its development-selected top-20 portfolios reach a $0.71$ median holdout Sharpe, selected on a split inside the optimization horizon. We report these single-run results descriptively, gross of costs, and are explicit about their limits throughout; in particular we do not isolate the effect of the leap machinery from the inherited search substrate, which we leave to future work.

Large Vision-Language Models (LVLMs) have achieved strong performance on diverse visual tasks, yet their ability to reconstruct and reason about the 3D structure of the scene depicted in 2D images – referred to as spatial intelligence – remains limited. Existing approaches attempt to address this gap by using real-scene spatial question answering datasets that require dense geometric annotations. However, constructing such labels is costly, time-consuming, and often noisy due to reliance on external perception modules. In this work, we propose a novel paradigm inspired by human cognitive development: learning foundational spatial skills through structured block-manipulation tasks. We introduce SpatialBlock-15k, a synthetic dataset of 15,000 block-stacking problems covering 3D-to-2D projection, viewpoint transformation, and structural combination. The dataset further incorporates controlled color modulation as visual cues to encourage anchor-based reasoning in visually complex conditions. Experiments demonstrate that LVLMs trained on our dataset through either direct answering or reasoning-based prediction significantly outperform baselines and generalize to real-world spatial tasks, despite the dataset's synthetic and compact nature. Code and data are available at https://github.com/rsoohyun/SpatialBlock.

Retrieval-augmented generation (RAG) critically depends on retrieving the evidence necessary for effective reasoning. However, this remains particularly challenging in multi-hop question answering (QA), where supporting passages are often linked through intermediate entities and relations that must be progressively uncovered. Existing retrieval approaches typically rely on a single retrieval intent or one-shot query expansion, limiting their ability to adapt to newly retrieved evidence and potentially introducing noisy or redundant retrieval signals. To address these limitations, we propose a training-free multi-hop retrieval framework that integrates evidence-conditioned exploration, passage-specific contrastive refinement, and coverage-aware final ranking. During offline indexing, the framework constructs passage-specific contrastive facets that characterize each passage relative to its semantically similar neighbors, providing fine-grained signals to distinguish closely related candidates. At inference time, the framework iteratively retrieves evidence, generates probes targeting unresolved information needs, refines candidate relevance using the contrastive facets, and selects a complementary set of passages that collectively cover diverse evidence-seeking intents. Experiments on MuSiQue, HotpotQA, and 2WikiMultihopQA demonstrate consistent improvements in retrieval quality and downstream QA performance over baselines.

Robotic manipulation often requires acting on information that is no longer visible, yet Vision-Language-Action policies are usually evaluated when the current observation largely determines the next action. Existing robotic memory benchmarks expose this gap, but they still rely mainly on final task success and therefore conflate forgetting with manipulation failure. We present MEMOBench, a benchmark for process level memory evaluation in robotic manipulation. MEMOBench includes 30 history dependent tasks, 1{,}500 expert demonstrations, and 4{,}200 executable checkpoint instances from 84 templates. Each checkpoint pairs coarse to fine language with a simulator predicate and labels one memory operation: Storage, Update, or Compression. These annotations define Memory Storage Rate, Memory Update Rate, and Memory Compression Rate, which measure memory fidelity alongside task success. Across standard and memory augmented VLA policies, the strongest memory module baseline reaches only 31.9% average success rate, and high storage often coexists with weak update and compression. Checkpoint language also supervises semantic, contrastive, and framewise memory alignment objectives, yielding modest gains across different memory operations. MEMOBench provides a diagnostic evaluation suite and training supervision for memory grounded robotic policies. The project page is available at https://github.com/Collab-Gen/MEMOBench.

We identify the Flow Moment, a reasoning pattern characterized by sustained, process-confirming verbalizations such as I'm doing, in contrast to the revision- and backtracking-oriented Aha Moment. We refer to their corresponding linguistic expressions as Flow Markers and Aha Markers, respectively. Based on this observation, we construct Flow-CoT by rewriting the discourse markers of original reasoning traces while preserving their underlying reasoning content, and use it as auxiliary supervision for on-policy self-distillation (OPSD). We further propose Aha-Flow Distillation (AFD), a dual-mode extension of OPSD that pairs different forms of privileged information with corresponding reasoning instructions. The Aha branch retains concise solution-based supervision, while the Flow branch introduces rewritten Flow-CoT under a direct and confident reasoning instruction. At inference time, the model uses only the standard reflective instruction, so Flow-style reasoning serves purely as a training signal. Experiments on AIME25 and HMMT25 show consistent improvements across Qwen3-8B and Qwen3-4B: AFD improves Avg@12 from 60.8 to 61.3 on Qwen3-8B and from 57.5 to 58.6 on Qwen3-4B over our reproduced OPSD baselines. Controlled ablations further show that, with the same Flow-CoT/Aha-CoT composition, dual-mode training improves Avg@12 from 59.5 to 60.1, indicating that the benefit comes not only from introducing heterogeneous reasoning supervision, but also from how it is organized during self-distillation. The code is available at https://github.com/Wang-Xiaodong1899/Aha-Flow-Distillation.

Frequency estimation is the problem of estimating individual tone frequencies from noisy multi-tone sinusoidal signals. Existing frequency estimation methods have difficulty accurately estimating both the number of tone frequencies and the individual tone frequencies in low signal-to-noise ratio (SNR) environments, because weak tone frequency components are buried in noise. In addition, existing methods generally exhibit a trade-off between robustness at low SNR and frequency estimation precision at high SNR, making it difficult to achieve consistently superior frequency estimation performance over a wide SNR range. To overcome these limitations, this paper proposes an SNR-adaptive frequency estimator (SAFE). SAFE consists of a time-frequency image neural network (TFINet), which enhances weak tone frequency components at low SNR, and an SNR-based frequency selector (SFS), which selects an appropriate frequency estimator according to the SNR of the estimated tone frequencies. TFINet enhances tone frequency components even in the low-SNR range, while SFS estimates the SNR of each tone frequency and selects either a robust frequency estimator or a super-resolution frequency estimator according to the estimated SNR. This enables SAFE to achieve robustness at low SNR while preserving high precision at high SNR. Simulation results show that SAFE achieves an False Negative Rate (FNR) of 13.00% over the SNR range from -10 dB to 0 dB, corresponding to an 13.04% improvement over the state-of-the-art method. In addition, SAFE reduces the Nearest Neighbor-Root Mean Squared Error (NN-RMSE) by 56.67% compared with the state-of-the-art method, demonstrating that SAFE performs more accurate frequency estimation. Furthermore, experiments using real-world data demonstrate that SAFE provides robust frequency estimation performance even in practical environments with clutter.

Localizing a font into new languages is a highly intricate task requiring precise design adaptation of glyphs, color/texture, and spacing/kerning, from source to target languages. Most existing methods focus on single glyph generation with limited capability in handling multilingual font rendering. In this work, we propose LoGAN, a VLM-based agentic framework for few-shot multilingual font localization, which takes in a small number of individual glyphs from a font or letters from a logo and uses them to generate complete character sets in other languages. LoGAN breaks down this task into multiple components: a glyph-level diffusion model, a style finetuning module, a spacing and kerning transfer algorithm, and a texture expansion model, with a VLM agent coordinator. LoGAN achieves broad language coverage for font localization with various styles, including Chinese/Korean/Japanese (CJK). We evaluate our approach on both font and real-world logo datasets spanning more than 27 languages and compare it against both specialized font generation and state-of-the-art image editing models with strong text rendering capabilities (e.g., FLUX, Nano-Banana). Our approach yields higher glyph fidelity while maintaining better style, texture, and kerning consistency according to both quantitative and qualitative evaluations.

Reasoning over privacy-constrained records requires combining structured attributes with evidence from free-text narratives. We introduce CIPHER (Cross-record Inference over Privacy-Hardened Evidence Records), a benchmark of expert-validated questions from consumer-finance, clinical, and law-enforcement records. The questions cover common tabular operations and include executable SQL supervision. We evaluate retrieval, prompting, table-specialist, and hybrid symbolic-neural systems under native redaction and surrogate-based evidence restoration. All system families exhibit substantial failures even when supporting records are provided. Most errors arise from incorrect record selection and predicate interpretation rather than arithmetic execution. Privacy transformations have non-uniform effects, sometimes obscuring necessary evidence and sometimes reducing distraction. CIPHER provides a reproducible testbed for diagnosing these failures and assessing how transformations of sensitive text affect reasoning over hybrid records.

Purpose: This study aims to develop an AI framework applicable for postoperative imaging for automated measurement of spinopelvic parameters on radiographs with robustness to the presence of spinal implants. Materials and Methods: We retrospectively reviewed lateral lumbar spine radiographs from two institutions (Internal: January 2017–December 2024; External: October 2021–September 2025). We developed the Restore, Segment, and Measure (RSM) framework, incorporating a novel Artifact Removal Network based on Autoencoding and Inpainting (ARNAI) to mitigate implant-related artifacts in postoperative radiographs. Segmentation and spinopelvic parameter (PT, LL, SS, SCA) measurement performance were assessed using Wilcoxon signed-rank tests and intraclass correlation coefficients. Results: When ARNAI was added to a recent Transformer-based segmentation model, FCBFormer, the mean DSC increased to 0.870 from 0.814, with marked gains at L3–L5 and smaller improvements at L1–L2. On 91 radiographs with implants, the mean L4–L5 segmental Cobb angle error decreased to 4.7 {\deg} from 15.6–16.2 {\deg}, an average error reduction of 70%. The ICC for L4–L5 segmental Cobb angle improved to 0.54 (Rater 1) and 0.59 (Rater 2) from 0.18, and ICCs for pelvic tilt, lumbar lordosis, and sacral slope all exceeded 0.70. The improvement in L4–L5 segmental Cobb angle error was statistically significant in the internal implant-containing cohort after correction for multiple comparisons. Conclusion: The proposed RSM framework improved automated spinopelvic parameter measurement in implant-containing postoperative radiographs. By mitigating implant-related artifacts, ARNAI improved segmentation and downstream measurement accuracy, with the greatest benefit observed for L4–L5 segmental Cobb angle estimation, where the mean error was reduced by approximately 70%.

Multi-head latent attention (MLA) exposes many logical query heads through one packed latent KV stream. This representation is memory efficient, but it removes the physical per-head cache boundary assumed by conventional head-wise reuse. We present our system, a DeepSeek-V4 realization of RedKnot's head-aware reuse principle. Each immutable document is processed offline at canonical position zero; certified Local-head contributions are retained as MLA-Off. At serving time, query-side RoPE relocation restores the document's request position, a small Global-head set and protected Local token rows are recomputed as MLA-Online, and the two paths are merged before a single shared output projection. The packed MLA latent is never split. DeepSeek-V4-Flash uses 37 reusable layers and a 56/8 Local/Global partition, giving a 75.29% analytic logical head-row ceiling; the Pro-0813 profile uses 55 layers and 112/16 heads, giving 78.89%. Frozen Flash operating points show hot-artifact TTFT speedups of 2.02-3.84x. At 256K, the archived three-dataset study reports an aggregate F1 change of +3.24 percentage points, an EM change of +4.16 points, and a 78.7-79.5% analytic major-operator arithmetic saving, while one dataset decreases by 2.81 F1 points. A separate author-reported 256K hot-artifact QPS measurement is approximately 2.0x; because its raw concurrency trace is not included in this bundle, we mark it as preliminary rather than archived evidence. We describe the factorization, position repair, token-row closure, sparse-MoE support, TP8 integration, and the measurement boundaries needed to interpret these results.

Recent text-to-audio-video (T2AV) models jointly generate video, speech, sound effects, and ambience from a single text prompt. This capability poses new challenges for safety evaluation, as unsafe content may be conveyed through the audio track or arise only when the visual and audio tracks are interpreted jointly. Existing safety benchmarks largely focus on either generated video or generated audio in isolation and are therefore not designed to capture these risks. To close this gap, we introduce AV-SafetyBench, the first safety benchmark developed specifically for T2AV generation. AV-SafetyBench comprises a four-axis, 13-category taxonomy and 5,200 manually reviewed prompts that specify visual scenes, speech, and non-speech audio. Our evaluation protocol assesses each output under three views: Full-AV, Video-Only, and Audio-Only. It then uses the Video-Only and Audio-Only judgments to assign Full-AV unsafe outputs to one of four risk sources: Video-Only, Audio-Only, AV-Both, or AV-Joint. We evaluate five open-source T2AV models and validate the automated Full-AV judgments against human annotations. Across the five models, Full-AV Unsafe Rates range from 25.1% to 49.4%. Beyond these aggregate rates, risk-source analysis reveals that, for four of the five models, Audio-Only and AV-Joint cases-unsafe outputs missed by video-only evaluation-account for 41.6-48.3% of Full-AV unsafe outputs for which a risk source could be assigned. In the Cross-Modal Harm Emergence category, AV-Joint accounts for 87.5% of unsafe outputs withan assigned risk source. Together, these findings demonstrate the value of AV-SafetyBench for evaluating T2AV safety across the visual and audio modalities and their interaction.

Generalized Nash Equilibrium Problems (GNEPs) often arise in multi-agent engineering applications that require distributed algorithms. Unlike traditional approaches that enforce consensus on multipliers, our method removes the need to share multipliers, reducing communication and improving privacy. As a result, different initializations can lead to different GNEs, including non-variational ones. We establish convergence under sufficient conditions using an input-to-state stability (ISS) framework.

Large language models achieve strong performance across diverse tasks, but deployment remains costly because of memory, latency, and energy demands. Structured pruning reduces these costs by removing architectural components, yet its recovery stage is often limited by a mismatch between the recovery module's representational capacity and the complexity of the removed knowledge. We call this bottleneck the capacity-knowledge asymmetry and propose OverRep, an Overcomplete Reparameterization framework for structured LLM pruning. Following the principle of "train overcomplete, deploy compact", OverRep temporarily overparameterizes the recovery module during training to absorb complex knowledge distilled from the original model. After recovery, the overcomplete re-parameterization is algebraically merged into a mathematically equivalent compact module, preserving the pruned model's inference-time architecture and computational cost. OverRep further introduces an annealed activation that enables nonlinear training dynamics while converging to a linear regime for exact algebraic merging. Across three backbone families, OverRep improves retained reasoning performance over strong recovery baselines by up to 5.5 and 8.4 points at 25% and 50% pruning, respectively, while keeping memory usage and TFLOPs comparable to existing recovery methods. Our code is available at https://github.com/mmai-laboratory/OverRep.

LLM agents complete tasks by issuing sequences of tool calls, and every observation they read is a channel through which an indirect prompt injection can enter. A successful injection has a characteristic shape when the trajectory is read in order: a benign prefix gives way to actions that serve the attacker rather than the user. Existing benchmarks measure whether such attacks succeed against live agents, and existing guard models judge a trace as a whole; no public corpus labels, step by step, where an injection enters a trajectory and which steps it corrupts. We present AgentDrift, a benchmark of 12,536 synthetic tool-call trajectories over five agent domains in which every one of the 71,024 steps carries one of four labels: benign, injection point, hijacked, or failed injection. The corpus contains 4,000 benign, 5,536 attacked, 1,500 failed-attack, and 1,500 hard-negative trajectories; attacked trajectories follow three compliance patterns whose label strings obey a stated regular grammar. Failed attacks carry an injection the agent resisted, and hard negatives carry legitimate content that resembles an attack, so a detector must separate attempt from success and deviation from novelty. Trajectories were generated by a single open model under category-specific protocols, enforced by a closed-vocabulary structural validator, screened by an LLM judge, and audited by hand on 1,200 trajectories; we show that the LLM judge was itself fooled by the hard negatives. A surface-feature logistic regression recovers only 55.4% of attacks (F1 0.647), including only 8.2% of partial hijacks and 23.1% of delayed executions, so nearly half of the attacks require modeling the behavioral sequence. We measure template concentration, attack-goal-family collapse, and world-identity leakage in the generated data, and release the corpus with its documentation under CC BY 4.0.

Ensuring effective transfer learning for vision-language models without compromising their generalization performance is crucial. However, many existing methods overlook data characteristics and simply reuse the training strategies adopted during pre-training. Specifically, they treat same-class samples as distinct instances and transform images independently of their paired text prompts, which makes model learning more difficult. We address these limitations through transformation-aware prompt conditioning and a re-calibrated contrastive loss. Fixed text descriptors identify the transformations applied to paired images, providing transformation-level consistency without altering class semantics. This design aligns the image and text branches at the transformation level, enabling richer representations while preserving the models' ability to generalize. In addition, our loss function mitigates positive-gradient dilution in soft-target cross-entropy when each anchor has multiple valid positives. During transfer, our approach treats same-class samples as positives rather than distinct instances, enabling the model to learn domain-specific features more effectively. Experiments across distribution shift, transfer learning, and few-shot settings demonstrate consistent improvements over existing approaches. Source code for our method is available at https://github.com/SoongE/ReCalCon.

Real-world remote sensing image dehazing (RSID) remains challenging because atmospheric scattering, spatially non-uniform haze and colour distortion jointly degrade structural and spectral information. Most deep learning methods rely on RGB inputs and spatial-domain feature extraction, which limits their ability to separate global background haze from local surface details. Here, we propose DPSF-Net, a dual-prior spatial-frequency network built on MCAF-Net for real-world RSID. The network uses hazy RGB images and dark channel prior (DCP) maps as joint inputs, allowing physical degradation cues to guide end-to-end feature learning. A spatial-frequency residual interaction block introduces a FourierUnit branch into multi-directional spatial interaction to model large-scale haze components. A prior-guided feature attention module adaptively fuses prior and attention features to reduce colour shift and structural distortion. A selective kernel complementary fusion module screens multi-scale skip features through bidirectional residual complementary gating and selective kernel fusion. Extensive experiments demonstrate that DPSF-Net achieves state-of-the-art performance on the real-world RRSHID remote sensing image dehazing benchmark and remains competitive across multiple synthetic datasets. Moreover, the proposed method strikes a favourable balance among restoration quality, parameter count and computational complexity, supporting the effectiveness of dual-prior spatial-frequency modelling.

Existing car-following prediction methods mainly optimize trajectory accuracy, while rarely considering whether predicted disturbances propagate realistically along a vehicle platoon. This limitation may lead to accurate but string-unstable predictions. We propose SSP-DMGTimeNet, a physics-constrained learning framework for spatiotemporal trajectory prediction of vehicle platoons. The model combines multi-scale temporal representations with cross-vehicle interaction features to capture complex and time-varying platoon dynamics. A propagation-delay-aware causal attention mechanism explicitly models upstream-to-downstream disturbance propagation by learning response delays between adjacent vehicles and accumulating them along the platoon. In addition, time- and frequency-domain string-stability losses relieve disturbance amplification across both adjacent vehicles and arbitrary sub-platoons during training. Experiments on HighD show that SSP-DMGTimeNet achieves an unstable-window rate of 0.65% for five-vehicle platoons and a maximum head-to-tail amplification of 0.898 on the ground-truth excitation subset, while maintaining competitive trajectory prediction performance. In zero-shot evaluation on NGSIM US-101 and I-80, the model achieves velocity MAEs of 1.316~m/s and 1.252~m/s, with unstable-window rates of 3.90% and 4.10%, respectively. These results demonstrate that incorporating platoon-level physical constraints can effectively balance trajectory prediction accuracy and disturbance propagation stability.

Invasive neural recordings provide high-fidelity measurements of brain activity, with signals such as intracranial EEG (iEEG) and intracortical spiking activity capturing neural dynamics at different spatial and temporal scales. Yet existing neural foundation models have largely been developed independently for different invasive recording paradigms, leaving joint pretraining across heterogeneous invasive signals underexplored. In this work, we introduce iBrain, a unified foundation model that jointly learns from iEEG and spiking activity. iBrain employs signal-specific encoders to accommodate their distinct signal characteristics and a shared spatiotemporal Transformer backbone to model dependencies across recording channels and time. We pretrain iBrain on over 7,000 hours of heterogeneous neural recordings using masked signal reconstruction and channel-view alignment, promoting contextual modeling of neural dynamics and robustness across different channels. iBrain consistently outperforms single-signal pretraining baselines and achieves state-of-the-art performance on multiple benchmarks. Further experiments demonstrate that iBrain exhibits transferability and data efficiency across diverse recording settings. These results highlight the potential of joint pretraining on heterogeneous invasive neural recordings to support scalable neural modeling and transferable representations across recording settings and downstream tasks.