Liuji Chen, Zeyu Zhang, Xinyuan Zhang, Shuai Nie, Qiang Liu, Shu Wu, Liang Wang
Personalized text generation requires models to capture user-specific writing styles from historical data. Existing approaches based on retrieval, parameter-efficient fine-tuning, or activation steering either introduce inference and storage overhead or struggle to separate stylistic signals from semantic content. We propose GLASS, a training-free framework for personalized generation via Global-Local Activation Steering with Sparse priors. GLASS uses sparse autoencoders to extract a global user-style prior from historical responses and constructs local contrastive style vectors over clustered interaction scenarios. During inference, it jointly injects global and local vectors into different model layers, enabling context-aware personalization without retrieval or parameter updates. Experiments on LaMP and LongLaMP show that GLASS outperforms retrieval-, fine-tuning-, and steering-based baselines across ROUGE metrics and LLM-as-judge evaluations. Further analyses show that SAE-based representations are more robust to topic and length shifts, suggesting better disentanglement of stylistic information from semantic residue.
AI Cosmos Team, Jinfeng Zhou, Yuxuan Chen, Fangming Chang, Mingyu Su, Yifan Zhu, Yuan Wang, Bo Zhang 等 35 位作者
Large language models (LLMs) can improve task performance during learning, yet these gains do not necessarily reflect learners' ability to thereafter apply the knowledge and skills independently. We present AI Cosmos, an agentic infrastructure for situated learning that organizes learning around two complementary principles: scenario-based participation and cumulative assets of practice. AI Cosmos organizes learners and role-specialized agents within learning scenarios that integrate task-relevant knowledge and, when needed, executable workbenches. An asset-memory loop preserves the processes and outcomes of practice as reusable assets with explicit provenance and dependencies. We evaluate scenario-based interaction through two preregistered randomized controlled studies examining the acquisition of legal reasoning and programming competence, holding the same underlying LLM across conditions. AI Cosmos achieves greater learning gains than a conventional LLM in legal reasoning and higher mastery of unseen programming problems than an LLM-as-Teacher baseline.
Long-horizon consistency is essential for video world models, yet it becomes difficult to maintain once observed content falls outside the context window. Existing memory mechanisms either retrieve history as implicit context, which remains constrained by context capacity and entangles static and dynamic information or consolidate history into explicit global representations, which struggle to model evolving dynamics and accumulate geometric errors. To overcome these limitations, we introduce \textbf{GROME}, a novel factorized memory framework that augments video world models for long-horizon coherence across both static scenes and dynamic entities. GROME operates via three synergistic components: 1)~\emph{Static-Dynamic Factorization} decouples historical evidence into static and dynamic components, preserving static scene geometry and dynamic entity states separately; 2)~\emph{Local Memory Retrieval} selects spatially aligned static history to eliminate geometric drift alongside the latest dynamic entity states to ensure temporal continuity; 3)~\emph{Training-Free Memory Injection} seamlessly grounds the factorized memory into video world models via their native conditioning pathways. Experiments on long-horizon revisit scenarios demonstrate that GROME substantially enhances scene persistence and dynamic entity fidelity over existing methods, all while preserving the visual quality and controllability of the frozen backbone without additional training. Beyond self-rollouts, GROME readily accommodates external observations as plug-and-play memory, unlocking flexible and controllable video continuation and scene exploration.
Yikun Zhang, Xiwei Cheng, Mingze Yuan, Yu Zhang, Brian Shim, Aarti Krishnan, Daniel Rosen, Kunyang Sun 等 18 位作者
Molecular discovery requires navigating vast combinatorial search spaces while balancing competing biological objectives. AI agents can automate expert-defined molecular design protocols, but automation alone does not address how to optimize workflows for new tasks or improve their development through experience. Here we introduce MoLRSI, an AI agent that combines systematic optimization of workflows and model architectures with persistent cross-task memory to recursively improve its search policy for molecular discovery. Workflow performance, refinement outcomes, and execution feedback guide subsequent searches, allowing the agent to improve both its workflows and its strategy for developing them. We evaluate MoLRSI in two challenging molecular discovery settings that require balancing competing properties. In antibiotic discovery, MoLRSI developed a screening workflow balancing antibacterial activity, cytotoxicity and structural novelty, achieving an experimental hit rate of 15.9%, compared with 8.7% for a previous expert-led workflow using the same training data and screening library. In de novo peptide-MHC binder design, MoLRSI developed workflows that jointly consider target binding and specificity, producing eight experimentally validated hits against MART-1 and SARS-CoV-1 pMHCs with preferential binding over tested off-targets while Claude Science completely failed. These findings demonstrate how experience-guided workflow optimization can address competing discovery objectives while improving the search strategies used in subsequent campaigns.
Self-evolving agent skills enable AI agents to accumulate and refine reusable procedural knowledge from interaction experience without updating model parameters. However, since such skills are typically represented as textual documents and executed through LLM reasoning, even previously learned procedures must be repeatedly interpreted and regenerated, introducing avoidable uncertainty into every execution. In contrast, program execution is deterministic and independently verifiable. Intuitively, transforming well-specified operations into executable programs can reduce reliance on LLM reasoning and, consequently, reduce execution uncertainty. Motivated by this intuition, we introduce SKillCompiler, a self-evolving skill framework that integrates programmatic compilation into the skill evolution process. Specifically, SkillCompiler adopts an executability-aware skill compilation strategy that decomposes skills into fine-grained operations, analyzes their executability, and constructs hybrid skill skeletons that retain LLM reasoning for context-dependent operations while compiling well-specified ones into validated programs. To avoid redundant program synthesis and enable executable knowledge to accumulate across evolving skills, we further design a program asset reuse mechanism that organizes validated programs as reusable assets, retrieves existing implementations whenever applicable, and synthesizes and verifies new programs only when necessary. Extensive experiments on five benchmarks and six model families demonstrate that SkillCompiler consistently outperforms state-of-the-art methods, achieving up to 32.33\% improvement in task performance.
Compositional motion transfer requires a model to extract motions from multiple reference videos, and apply each motion to a user-specified object in one target scene. Existing methods normally struggle to reliably assign different reference motions to the correct target objects, thus a motion intended for one object may affect another. Meanwhile, some motion-specific tuning approaches are still facing another issue: they must run a new optimization whenever the reference video changes, making repeated transfers inefficient. To address these limitations, we propose RART, a framework that unifies object-wise motion extraction, target-region anchoring, and frame-aligned injection into a frozen video Diffusion Transformer for multi-object motion transfer. For each reference motion, RART extracts temporally ordered features from the masked reference video, and then conditions them on the target first-frame mask to form mask-anchored motion tokens. Frame-aligned motion cross-attention allows each generated frame to read only the motion tokens at the corresponding time slots, composing all object motions in one denoising process. During training, a motion-attention routing loss encourages each target region to attend to its corresponding motion tokens and reduces interference between objects. Therefore, RART can handle novel reference-target combinations without per-video optimization. Experiments show that RART achieves accurate motion preservation, efficient inference, and strong compositional generalization.
Yi Wen, Xiangjue Dong, Bingxuan Li, Hangxiao Zhu, Leo Zhicheng Liu, Yu Zhang, James Caverlee
Deep research agents (DRAs) conduct long-horizon reasoning, search, and literature synthesis to provide long-form answers to complex, domain-specific questions. These agents are typically trained and evaluated on fully specified user queries. However, real-world user queries can often be underspecified. We introduce UNDERSPEC-DR, a framework for systematically evaluating how underspecified queries affect agents' final answer quality and intermediate search behavior. Across three DRAs we evaluated, underspecification substantially reduces how well DRA's answers fulfill users' information needs, even though they remain grounded in cited evidence. Our analysis of intermediate search behavior suggests a possible explanation: DRAs conduct broader searches that draw less on the original user query but rely more on their learned prior during training, which can be misaligned with users' intent. Our findings reveal the vulnerability of current DRAs to underspecified user queries and inform future efforts to build DRAs that proactively recognize when additional user input is needed to reliably complete deep research tasks.
Developing seamless, high-performance, native intelligent full-duplex Spoken Language Models (SLMs) remains a critical challenge and long-standing goal for the speech and NLP community. Despite notable progress, recent endeavors are fundamentally constrained by severe modality interference, which causes substantial knowledge degradation and compromises semantic integrity -- ultimately making full-duplex SLMs feel unnatural and unintelligent. In this paper, through an exhaustive fine-grained analysis of model optimization dynamics, we uncover the root cause of such performance degradation, revealing that modality interference arises from inherent gradient conflicts between acoustic and semantic modeling when the two modalities are forced to share a deep parameter space. Guided by this key insight, we introduce Lychee-FD, a native end-to-end full-duplex framework designed to mitigate modality interference. Importantly, we propose a hierarchical parameter separation strategy that decouples conflicting modalities in deep layers while preserving cross-modality coherence via a dedicated semantic alignment channel. Extensive experiments on multiple full-duplex benchmarks demonstrate that our method significantly advances the state of the art, yielding substantial improvements in both speech intelligence (+7.4% on Spoken QA) and full-duplex interaction fluidity (+28.5% on FullDuplexBench 1.5) without compromising inference efficiency. To the best of our knowledge, this work is the first to achieve two key advances: 1) uncovering and elucidating the root cause of modality interference in full-duplex SLMs, and 2) designing an elegant hierarchical model together with a practical solution for seamless, high-performance, native intelligent full-duplex SLMs.
Semantic entropy, the entropy of the meanings among a language model's sampled answers, is estimated from the verdicts of a judge that declares pairs of answers equivalent or not, and automatic judges err in ways that depend on the answers they compare. The existing minimax theory covers only judges whose false-equivalence probability is the same for all pairs. We model a known judge that declares two answers with meanings $a$ and $b$ equivalent with probability $q+ΔM_{ab}$, for a symmetric matrix $M$ and $q$ bounded away from $0$ and $1$, and show that for three meanings the rate is set by tree statistics on four answers, while on an open set of channels the consistency threshold is set by trees on five. For every channel in an explicit open class, the minimax squared risk is of order $N^{-1/3}$ with $N=n^4Δ^6$ on polynomial weak windows on which $n^5Δ^8$ exceeds a multiple of $\log n$, and of order $\min\{1,N^{-1/3}\}$ on every polynomial weak window when the tree functions on at most four vertices are injective. The exponent comes from a cubic degeneracy at the equitable point. On an open set of channels in the class, the four-vertex tree functions take equal values at two distributions with different entropies. For such a channel, on every polynomial weak window, the minimax risk is of the order of the sum of $\min\{1,N^{-1/3}\}$ and a term that decays exponentially in $n^5Δ^8$, so no estimator is consistent while $n^5Δ^8$ stays bounded, even as $N\to\infty$. For $Δ=n^{-β}$ with $β>1/2$, channels of the class with injective four-vertex tree functions allow consistent estimation exactly for $β<2/3$, and the channels of this open set exactly for $β<5/8$, at which the risk exponent drops from $1/12$ to $0$. Computer-assisted certificates show that both kinds of channels occur on open sets.
Reinforcement learning (RL) has substantially improved the ability of large language model (LLM) agents to interact with environments and solve multi-turn tasks. However, effective agentic RL remains challenging: sparse outcome-only rewards provide limited guidance for assigning credit to individual steps within long interaction trajectories. Existing approaches often introduce dense intermediate supervision, such as process reward models or auxiliary self-supervised signals, which increases supervision and tuning complexity and may limit generalization across tasks and domains. We present AEM, a supervision-free credit assignment method that adaptively modulates entropy dynamics during RL training to improve the exploration-exploitation trade-off. Since in agentic RL the environment is typically affected by a complete response, rather than an individual token, our analysis lifts entropy dynamics from the token level to the response level, aligning uncertainty estimation with the effective action granularity of LLM agents and reducing sensitivity to token-level sampling noise. We further show that entropy drift under natural-gradient updates is governed by the interaction between the sampled-response advantage and its relative surprisal. Motivated by this result, AEM derives a practical response-level uncertainty proxy and uses it to rescale advantages, leveraging the evolving balance between positive and negative samples to naturally transition from exploration to exploitation. Extensive experiments on ALFWorld, WebShop, and SWE-bench-Verified with models ranging from 1.5B to 32B demonstrate that AEM consistently improves strong RL baselines, including a +1.4\% gain when integrated into a state-of-the-art software-engineering RL training framework.
Pneumonia changes across respiration, inflammation, and organ function, while the measurements that describe it arrive at different times and are often incomplete. We introduce PULMOWORLD, a world model for longitudinal pneumonia dynamics that connects asynchronous state estimation, recursive simulation, and evidence-constrained planning. History-aware filtering separates state propagation from observation-driven correction, so a measurement updates the shared belief only when it is observed. That belief supports recursive forecasts of clinical values and modality representations, and summaries of future prognosis. Cap-MBRL searches imagined continuations inside an authorized action set and replaces a frozen fallback only when a candidate meets the risk budget and a predicted-cost margin. On the pneumonia cohort MASS, recursive forecasts are competitive with task-trained sequence models, and simulated future states retain prognostic information. Trained separately on public MIMIC-IV-CXR chest radiographs and clinical trajectories, this model improves prognostic discrimination over latest-value fusion in development and is competitive on long-horizon error on the locked pneumonia panel. In five known-mechanism MASS-Syn environments, Cap-MBRL lowers selection regret relative to direct fallback execution.
Adrien Petralia, Alexandre Gavaudan, Artjom Joosen, Etienne Goffinet, Richard Kurle, Yassine El Ouahidi, Yassir Bendou, Abdelaziz Bounhar 等 356 位作者
We introduce Timestral, a family of four time series foundation models for probabilistic forecasting, with 5M, 28M, 123M, and 209M parameters. Timestral sets a new state of the art on GIFT-Eval, TIME, and fev-bench, where Timestral-209M and Timestral-123M take the top two places on every metric. We achieve this with three ingredients: Decoupled Horizon Forecasting (DHF), which separates prediction tokens from the causally encoded context, so that longer horizons are trained without masking context and each observed value is both a forecast target and context for later forecasts; a covariate framework that natively supports known-future covariates within a causal architecture; and a two-stage pipeline that pretrains exclusively on a diverse, large-scale synthetic corpus before post-training on a mix of real and synthetic data with extended context. This report describes the model design and its benchmark results. Weights for all four models are available on Hugging Face and inference code on GitHub.
Diffusion video models produce striking clips, but they cannot reliably render an exact formula, a correct chart, a rule-consistent game or narration synchronized to on-screen events. A different route is emerging: multimodal agents that write code to render every frame and synthesize every sound. We introduce VibeVideoBench, a benchmark for evaluating agents on code-driven video generation. Its 239 self-contained tasks in English and Chinese span 12 categories, from data stories and music-driven pieces to classroom lessons and paper explainers; each asks an agent, running in its native harness inside a sandbox with no image or video generation model, to deliver a finished MP4 with sound within one hour. Evaluation is hierarchical and evidence-based: program checks of the delivery spec, about 13k task-specific rubric checks traced to the requirements they enforce, and five quality dimensions anchored per task, scored by three VLM judge panels. Across nine frontier agent systems and 2,151 scored rollouts, GPT-6 Astra and Claude Opus 5.5 rank first and second with both the judges and human annotators, and the strongest system satisfies every core requirement on 76.3% of tasks. Delivery is a second axis: counting missing videos as zero changes the ranking, and three quarters of the lost videos are final renders that started too late, ran too slowly or never started, a failure of how agents plan their hour. Agents build their own rendering pipelines and inspect their own frames, yet how much they look does not predict quality. The judges' ranking of the systems closely matches blind human preferences.
Mixture-of-Experts (MoE) inference on memory-constrained edge platforms is often dominated by expert-weight movement rather than computation. Although each token activates only a small subset of experts, routing is input-dependent, so all experts must remain accessible and memory movement dominates latency. Existing approaches such as uniform pruning and expert dropping overlook two properties of MoE execution: expert combinations are highly skewed in frequency, and experts contribute asymmetrically within each combination. We present CRISP, a combination-aware framework that co-designs pruning and runtime scheduling for efficient MoE serving on memory-constrained edge platforms. CRISP reorders neurons by importance so one stored expert can support multiple execution widths, assigns differentiated execution to hot, warm, and cold combinations based on frequency and expert role, and exploits heterogeneous widths to overlap expert loading with computation. Across Mixtral-8x7B and DeepSeek-MoE-16B, CRISP improves throughput by up to $2.1\times$ at the same memory budget while limiting accuracy loss to 3.6\% relative to full-width inference. Our results highlight that efficient edge MoE serving requires jointly optimizing which expert capacity is preserved and how expert execution is scheduled.
Functional segmentation requires locating small interactive parts, such as knobs, buttons, and sockets, in indoor point clouds and returning confidence-ranked instance masks. A point-wise semantic predictor produces a continuous probability field, whereas the benchmark evaluates a discrete set of scored instances. This creates a substantial field-to-instance gap. We present TopoField, a deterministic decoder built on top of VOLT. TopoField first converts 13 test-time forward passes into a semantic consensus field, then decodes instance candidates from the consensus geometry, and finally ranks the candidates in a class-aware manner. The augmentation stage retains the processed-point identity of every fragment prediction, accumulates probabilities in a common point array, normalizes by coverage count, and fuses the resulting caches. The decoding stage separates high-confidence cores from broader support regions, constructs a symmetric radius-pruned point graph, and uses a multi-source graph-geodesic partition to split support regions that contain multiple seeds. On 561 ground-truth instances across 30 validation scenes, an IoU-0.5-valid candidate exists for 67.2% of instances, and top-1 coverage is 20.0%. These complementary measurements motivate joint attention to candidate construction and ranking. TopoField achieves 28.286% AP, 41.383% AP50, and 47.558% AP25. TopoField achieved second place on the ECCV 2026 SceneFun3D Challenge.
This report presents the implementation and validation of two end-to-end multimodal adaptation pipelines for 3D-aware visual question answering on a custom spatial dataset. The first pipeline fine-tunes HiSpatial-3B, which augments RGB observations with metric-scale 3D point maps; the second adapts Qwen3-VL through memory-efficient QLoRA-based supervised fine-tuning. The completed engineering stack covers dataset conversion, image-annotation alignment, monocular geometry preprocessing, instruction construction, mixed-domain training, checkpoint management, batch inference, and automated evaluation. On the target spatial benchmark, HiSpatial adaptation increases overall accuracy from 1.73% before fine-tuning to 79.89% at 3,600 steps. On a separately collected styled split, accuracy improves from 14.9% to 55.1%, indicating transfer beyond the original training set. The Qwen3-VL pipeline is likewise implemented end-to-end for both 8B- and 32B-scale models using mixed-domain QLoRA adaptation. The resulting accuracy improves from 20.74% to 45.27% for Qwen3-VL-8B and from 30.79% to 47.60% for Qwen3-VL-32B. Together, these results establish two complementary and reproducible adaptation pipelines for geometry-aware and general-purpose multimodal spatial reasoning. These results establish a reproducible engineering foundation for subsequent work on metric 3D reasoning, spatial grounding, and domain-adapted multimodal systems, while also identifying training-data composition as a central factor in preserving general multimodal capability during specialization.
Large language model (LLM) agents rely on memory systems to leverage long-term interactions, yet existing systems typically retrieve memories based on their relevance to the current query and use them as contextual evidence. However, relevance alone does not characterize how a memory contributes to task solving: the same memory may serve different roles across tasks, such as directly supporting the answer or linking to complementary memories with weaker semantic relevance but necessary information. We therefore reframe agent memory systems as Task-Oriented Role-Adaptive Workspaces, where memories are selected and organized by their task-oriented roles and relations. Based on this view, we propose TrawMem, which organizes interaction histories into local threads and constructs task-oriented workspaces by assigning roles and jointly selecting complementary memories. Experiments on LoCoMo and MemGallery with four open- and closed-source backbones show consistent improvements over representative memory systems, with macro-F1 gains of 2.23-3.14 points. On LoCoMo, TrawMem improves the LLM-Judge score over SimpleMem by 4.07 and 4.31 percentage points on Qwen3.5-9B and DeepSeek-V4-Flash, respectively, while requiring approximately 12.8% and 5.1% fewer context tokens, respectively.
Jiaming Hu, Shi Bo, Yan Zheng, Tian Wang, Florian Dubost, Alejandro Mottini, Junze Liu, Arvind Srinivasan 等 14 位作者
Joint-embedding predictive architectures (JEPAs) enable world models to plan from images by predicting future embeddings. However, their visual prediction objectives do not explicitly incorporate task-relevant state information, so the physical state that determines task progress may not be reliably captured in the latent space. Supervising world models with simulator state can make planning more informative, but it requires deciding which state features matter. Features are typically selected by hand or weighted uniformly, and a uniform weighting gives every standardized state feature equal influence, so redundant coordinates can overcount the same variation, while irrelevant ones can obscure task-relevant differences. To address this limitation, we propose SWIM (State Weighting with Inverse-Dynamics Model), which learns state weights from offline trajectories by predicting recorded actions from current states and observed future displacements. The learned weights quantify each feature's contribution to action prediction and define a weighted state metric for state-supervised training, which we demonstrate by aligning latent and state distances. The inverse-dynamics model is used only during training, so the resulting world model plans from images with no additional planning-time computation. Across five simulated robotic manipulation tasks, SWIM consistently improves planning success over uniform state weighting.
Hanbin Hong, Shuang Wu, Shuya Feng, Nima Naderloui, Shenao Yan, Jingyu Zhang, Ali Arastehfard, Heqing Huang 等 9 位作者
Prompt-security claims for large language models (LLMs) are often compared as properties of individual attacks or defenses, although measured harmfulness depends on the target model, access setting, attack and defense configuration, prompt sample, and judger. We systematize 272 studies through three linked, multi-label taxonomies of jailbreak attacks, defenses, and model vulnerabilities. To make these dependencies explicit, we introduce an access-compatible evaluation protocol over configured $\langle M,A,D,S,J\rangle$ tuples and instantiate it in PromptSecurity, an open-source platform that records method settings, outputs, judgments, and query costs. We release two complementary datasets: JailbreakDB, with 445,752 jailbreak system-user prompt pairs and 1,094,122 benign prompts, and PromptSecurity-Eval, with 166,200 harmful-query and 25,800 benign-utility output records. We evaluate 11 target models, 20 attacks, nine defenses, and three automated judgers under a matched, single-turn protocol using 100 harmful queries and separate benign-utility tasks. Lower attack success can reflect failure to execute transformed requests rather than stronger safety; attack-category advantages can reverse when one method is excluded, and attack success varies with assistant choice. Defenses can reduce harmful outputs while degrading benign-task utility or causing backfire. Rejudging identical outputs also shows that measured attack success is sensitive to the behavior supplied to the judger, highlighting the need to report judger configurations alongside ASR. These findings motivate configuration-aware reporting rather than context-free method rankings. Resources: https://datasec-lab.github.io/PromptSecurityLeaderboard/ https://huggingface.co/datasets/youbin2014/JailbreakDB https://huggingface.co/datasets/youbin2014/PromptSecurity-Eval https://github.com/datasec-lab/PromptSecurity
Jianwen Chen, Peng Xia, Kaide Zeng, Siwei Han, Zhen-Yu Zhang, Charles Fleming, Xinyu Ye, Huaxiu Yao
Large language model (LLM) agents interact with their environments through executable harnesses that govern tool use and feedback. Optimizing a harness for a fixed policy overlooks how its behavior changes during learning, while training a policy under a fixed harness leaves the interface that generates its training experience unchanged. This motivates adapting the harness throughout policy training. We introduce CoHARL, a population-based reinforcement learning framework that co-evolves a shared policy and a population of executable harnesses. The policy learns from rollouts across the population, and a proposer uses its execution traces to revise harness code during training. CoHARL achieves a 27.6% mean resolved rate on SWE-rebench, improving over harness-only search by 6.4 percentage points. Experiments on SWE-bench Verified and Terminal-Bench 2.0 show gains in software engineering and terminal-based tasks. Under matched training budgets, CoHARL also outperforms searching for a harness once before RL and alternating policy updates with revisions to a single active harness.
Bin Wang, Yan Zhong, Zhengyu Zhang, Liang Luo, Buyun Zhang, Ellie Wen
Language models now propose hypotheses faster than anyone can check them, so the step that certifies them must need nothing from the proposer but the hypothesis. We state three requirements for such a step -- per-hypothesis automatable, attributable and diagnostic -- and introduce NxN e-valuation, which meets them. A hypothesis names a scope, the units it is about, and gives each of them a restriction that scores its outcome. The units of one scope serve as nulls for one another, so no resampling model is written per hypothesis; the algorithm certifies a stable, scope-level association and reports whether the scope, the restriction, or which unit receives which restriction carries the prediction. For independently drawn units, betting e-values bound the chance of a false certificate for each hypothesis in finite samples, under stated nulls. A betting lower bound on the share of units that carry the association gates every certificate, and an asymptotically valid equivalence test calls the restrictions shared only when it rejects differences between them above a declared noise ceiling. We test it on planted data, on randomized intervention data, and on real purchase data with 285 hypotheses that language models wrote without knowing our method or our classes, labeled by human experts. Meeting all three requirements costs no accuracy: NxN makes 692 correct decisions in 693 with one false certificate, against 659 with four for a composite of CRT tests that returns our verdicts but needs a resampling model per hypothesis, and 224 with 468 for held-out validation. Removing any of our statistics merges a verdict that certifies with one that rejects. The certificates concern predictive association, not causality.
Vision-Language-Action (VLA) models remain brittle in long-horizon, contact-rich manipulation because success-only imitation provides little supervision for execution drift, while failed rollouts are often discarded. We introduce RePO-VLA, a recovery-driven policy optimization framework that assigns distinct roles to success, recovery, and failure trajectories. RePO-VLA first applies Recovery-Aware Initialization (RAI), slicing each corrective suffix into a new episode so corrective actions depend on the current adverse state rather than the preceding failure. It then learns a Progress-Aware Semantic Value Function (PAS-VF), aligning spatiotemporal trajectory features with instructions and successful references. The resulting labels salvage useful failure prefixes via reliability decay, while low-value labels mark drift and terminal breakdowns, teaching differences among nominal, failed, and corrective actions. The data engine turns adverse states into planner-generated or human-collected corrective rollouts, teaching recovery to the success manifold. Value-Conditioned Refinement (VCR) trains the policy to prefer high-progress actions. At deployment, a fixed high-value token biases actions toward the learned success manifold without online failure detectors or heuristic retries. We introduce the Failure-Recovery Benchmark (FRBench), with standardized error injection and recovery-focused evaluation. Across simulated and real-world bimanual tasks, RePO-VLA improves robustness, raising adversarial success from 20% to 75% on average and up to 80% in scaled real-world trials. Code and video results are available at https://1018weijia.github.io/RePO-VLA-Web/.
Few-step distillation is essential for deploying modern video diffusion models. Among existing approaches, Distribution Matching Distillation (DMD) has emerged as a highly promising paradigm by achieving strong output-space fidelity without adversarial training, while remaining flexible across various student architectures. We observe, however, that DMD fundamentally relies on an output-space objective, lacking explicit regularization on the model's internal representations. As the distillation step count aggressively shrinks, this output-only supervision becomes insufficient, causing the distilled student to degrade in both fine-grained spatial structures and coherent temporal dynamics. To address this limitation, we introduce Self Flow Forcing (SF$^2$), which incorporates a predictive representation loss into the DMD framework without altering its output objective. By providing the student and an Exponential Moving Average (EMA) teacher with heterogeneous noise levels, we establish an information asymmetry. This formulation forces the student to predict the teacher's internal representations from a more corrupted input, effectively integrating self-supervised representation alignment into distribution matching. Extensive experiments demonstrate that SF$^2$ effectively mitigates temporal flickering and consistently improves both spatial fidelity and motion quality. Furthermore, the proposed mechanism is architecture-agnostic - seamlessly supporting both bidirectional and causal models - and requires no external feature extractors, introducing minimal computational overhead.
Jianyi Hu, Hangtao Zhang, Yi Liu, Yeqi Zeng, Li Zeng, Xianlong Wang, Rui Wang, Leo Yu Zhang
Typed decision models (TDMs), such as Jev, answer a question about an input with a probability, an option, or an ordinal level. Web services execute these decisions automatically above a confidence threshold. An attacker who writes part of the input can change a decision that the service then executes. Current adversarial benchmarks cannot measure this risk, because they judge generated text. A TDM writes no text, and its wrong decisions have a valid format. We present JevAdvBench, to our knowledge the first unified benchmark for the adversarial robustness of TDMs. It contains 862 questions with audited answers and twelve black-box attack methods that edit the input or the options by adaptive search. Because a changed decision is not always wrong, a review keeps only 5,586 answer-preserving variants, and we attack only correct decisions. The attacks flip 46.3% of Jev's correct decisions, while identical re-runs flip only 0.17%. A flawed argument without a command flips 77.2%, close to direct prompt injection at 92.0%. The same variants flip 34.0% to 51.7% of correct decisions on eight open TDMs. At a 0.8 threshold, 23.5% of the decisions that the service still executes are wrong. On Agentgateway, the Jev guardrail rejects a jailbreak request but accepts it with an inserted argument. Web services should therefore treat TDM inputs as untrusted and should not act on confidence alone. Project website: https://JevAdvBench.github.io/JevAdvBench/
Reinforcement learning with verifiable rewards (RLVR) has delivered impressive gains in mathematical and multimodal reasoning and has become a standard post-training paradigm for contemporary language and vision-language models. However, the RLVR recipe introduces a significant risk of capability regression, in which models forget foundational skills after prolonged training without employing regularization strategies. We empirically confirm this concern, observing that open-source reasoning models suffer performance degradation on core capabilities such as perception and faithfulness. While imposing regularization terms like KL divergence can help prevent deviation from the base model, these terms are computed on the current task and therefore do not guarantee preservation of broader knowledge. Meanwhile, commonly used experience replay across heterogeneous domains makes it nontrivial to decide how much training emphasis each objective should receive. To address this, we propose RECAP-a replay strategy with dynamic objective reweighting for general knowledge preservation. Our reweighting mechanism adapts online using short-horizon signals of convergence and instability, shifting the post-training focus away from saturated objectives and toward underperforming or volatile ones. Our method is end-to-end and readily applicable to existing RLVR pipelines without training additional models or heavy tuning. Extensive experiments on benchmarks using Qwen2.5-VL-3B and Qwen2.5-VL-7B demonstrate the effectiveness of our method, which not only preserves general capabilities but also improves reasoning by enabling more flexible trade-offs among in-task rewards.
Large language models incur substantial training and inference costs in personalized search and recommendation, as growing interaction histories lengthen request inputs and repeatedly expose the same item text to the model. We propose \textbf{C}omposable and \textbf{R}eusable \textbf{I}tem \textbf{M}emory (CRIM), a framework that separates item-level semantic encoding from request-level history modeling. A small set of learnable summary tokens compresses each item into compact continuous memories, while segment-level readout tokens (REC tokens) introduce predictive supervision within the interaction history. These memories are independent of users and queries and can therefore be reused across requests. At inference time, CRIM composes cached memories in interaction order and incorporates an optional query to produce a personalized retrieval representation, reducing repeated item-text encoding while retaining request-dependent contextual modeling. We further introduce a joint training mechanism that combines cached memories with sampled live re-encoding. Re-encoding only a few history items per request allows the retrieval objective to jointly optimize the item and request encoders at limited computational cost, while periodic cache refreshes maintain consistency. Extensive evaluation and cost analysis on MovieLens-1M, Qilin, and our new LazSeq dataset show that CRIM achieves state-of-the-art retrieval performance with substantially lower inference costs at the same backbone scale, while limiting item re-encoding during joint training. We will publicly release our code and dataset.
Multi-Turn Long-Form Question Answering (MT-LFQA) is a key application paradigm of Large Language Models (LLMs) in knowledge-intensive domains. However, existing benchmarks are limited to single-turn dialogue, while multi-turn dialogue benchmarks typically assess other orthogonal capabilities rather than knowledge-intensive factuality. To bridge this critical gap, we introduce \textbf{KnowMT-Bench}, a new benchmark designed to systematically evaluate MT-LFQA for LLMs across knowledge-intensive fields, including medicine, finance, and law. To faithfully assess the model's real-world performance, KnowMT-Bench employs a dynamic evaluation setting where models generate their own multi-turn dialogue histories given logically progressive question sequences. Our experiments on a diverse suite of LLMs show a clear degradation in both factual capability and information delivery efficiency in multi-turn dialogue. We further probe the underlying causes and investigate mitigation strategies spanning domain-specific finetuning, active context denoising, and RAG. The results show domain-dependent gains from finetuning, while active context denoising and RAG more consistently mitigate these issues, with RAG notably reversing this performance degradation. These findings underscore the importance of our benchmark for evaluating LLMs' conversational factual capabilities in real-world applications. Code and data are publicly available.\footnote{\href{https://github.com/Sven-LI-sankyuu/KnowMT-Bench}{\texttt{https://github.com/Sven-LI-sankyuu/KnowMT-Bench}}}
Jinyu Zhang, Lijun Han, Feng Jian, Lingxi Zhang, Hesheng Wang
In mobile robot shared control, effectively understanding human motion intention is critical for seamless human-robot collaboration. This paper presents a novel shared control framework featuring planning-level intention prediction. A path replanning algorithm is designed to adjust the robot's desired trajectory according to inferred human intentions. To represent future motion intentions, we introduce the concept of an intention domain, which serves as a constraint for path replanning. The intention-domain prediction and path replanning problems are jointly formulated as a Markov Decision Process and solved through deep reinforcement learning. In addition, a Voronoi-based human trajectory generation algorithm is developed, allowing the model to be trained entirely in simulation without human participation or demonstration data. Extensive simulations and real-world user studies demonstrate that the proposed method significantly reduces operator workload and enhances safety, without compromising task efficiency compared with existing assistive teleoperation approaches.
Tianyu Zhan, Gui Ling, Ran Tao, Kunhai Lin, Tong Xiong, Wang Yang, Kaixuan Zhang, Zhihong Chen 等 12 位作者
Generative retrieval (GR) typically relies on item-level beam search to retrieve candidates, leaving cross-item dependencies unmodeled and limiting direct optimization of list-level objectives. A natural alternative is to generate the entire list within a single autoregressive trajectory. However, greedy decoding with directly trained list-wise models substantially underperforms beam search. We identify two contributing factors: the broader search space explored by beam search and the positional bias of scarce positive signals, which deprives later positions of adequate supervision. To address these challenges, we propose $\textbf{SLIDE}$, a search-to-list item distillation framework that decouples offline search from online list generation. A teacher model retrieves high-quality candidates through offline item-level beam search, providing additional conversion-relevant supervision beyond sparse behavioral labels. Their Semantic IDs (SIDs) are then concatenated in beam-score order to form list-wise distillation targets. This search-to-list conversion transforms candidates retrieved along separate search paths into sequential item-generation targets within a single trajectory. Trained on these targets, the student generates the Top-$K$ list in a single greedy pass, also decoupling teacher scaling from online deployment. Experiments on FORGE and Taobao Search show that SLIDE substantially improves retrieval quality under single greedy decoding and, with stronger offline teachers, surpasses the beam-search reference. Supplementary RL experiments support the effectiveness of list-level generation.
After each Web search, a search-augmented model must decide whether its evidence suffices: answer, search again, or abstain. Prevailing systems leave this to holistic model judgement, which can accept a context naming every entity yet lacking one required fact. We instead make sufficiency structural: a per-query requirement hypergraph states which facts jointly yield another fact. This raises two challenges: deriving requirements from the question alone, and linking each requirement to passages about the right entity. Certify therefore decomposes the question, searches per requirement, and accepts links only after entity checks. Horn closure over the links yields a certificate, the minimal sufficient passages and their derivation, or an abstention naming open requirements. Given correct retrieval, requirements and links, the decision is sound, and complete up to the entity check. Pooled over three corpora with unanswerable twins, Certify decides significantly more accurately than search agents, graph retrieval and sufficiency judges in default configurations. On the same evidence, it matches a proprietary model's sufficiency label without an additional model call. Over 21M Wikipedia passages, entity checks add 4 to 13 points of decision accuracy where gold evidence exists. Finally, answers from its certificate are 80% correct, against 57% from all retrieved passages.