Large language models can solve many multi-step reasoning problems, but their reasoning remains brittle: plausible intermediate steps often lead to incorrect answers. Improving reasoning at test time requires balancing two imperfect signals: the generator, which provides diversity, and the verifier, which guides correctness. Existing approaches largely favor one side of this tradeoff. Verifier-weighted voting relies on trajectories sampled from the generator and cannot steer generation toward correct ones. By contrast, beam-style verifier-guided search can steer generation, but may over-concentrate on narrow trajectory sets and amplify verifier errors. We propose Verifier-Guided Self-Consistency (VGSC), a stochastic decoding framework that combines step-level correctness and future-outcome estimates to guide reasoning during generation. VGSC repeatedly proposes candidate next steps from the base language model and uses both signals in an acceptance/resampling procedure, while maintaining multiple trajectories for final answer-level self-consistency. We provide a probabilistic interpretation of VGSC as approximate sampling from a verifier-tilted trajectory distribution. Across multiple reasoning benchmarks, generator models, and verifier families, VGSC achieves higher average accuracy than representative baselines.
Embodied and edge systems increasingly rely on vision-language models (VLMs) to support heterogeneous skills, yet deploying a separate task adapter for each skill while retaining a high-precision backbone creates substantial memory and management costs. We present UniCompress, a deployment-oriented framework that converts multi-task expertise into one compact, adapter-free model. Starting from a consolidated BF16 Qwen3-VL-4B teacher, UniCompress constructs an asymmetric INT2 student using GPTQ initialization and teacher-guided quantization recovery. The approach jointly quantizes and recovers the projection weights in the vision encoder and language backbone, enabling aggressive compression of the complete vision-language model while preserving its multi-task behavior. We evaluate UniCompress on four heterogeneous embodied tasks: posture recognition, household-safety understanding, multimodal dialogue routing, and mobile-GUI action prediction. UniCompress reduces the approximate weight storage from 8.9 GB in BF16 to 2.1 GB in INT2, corresponding to a reduction of approximately 76%. It improves the INT2 macro task score by 34.3 percentage points over GPTQ-INT2 while remaining within 0.9 points of the consolidated BF16 teacher. These results demonstrate the potential of teacher-guided joint vision-language INT2 recovery for deploying a single multi-skill VLM on resource-constrained embodied platforms.
Yan Lin, Jilin Hu, N. M. Anoop Krishnan, Morten M. Smedskjaer
Amorphous (disordered) materials require simulation cells with at least hundreds to thousands of atoms to accurately capture their diverse local atomic environments, compared to crystalline materials that can be described by unit cells containing a few to hundreds of atoms. To advance the design of amorphous materials with desired properties and facilitate the exploration of their vast design space, generative inverse design has emerged as a promising approach. It aims to directly output materials with properties closely aligned with the desired ones using probabilistic generative models conditioned on desired properties, which can be more resource efficient than the traditional trial-and-error approach. However, due to the inherent stochasticity of probabilistic generative models, when element assignments are unconstrained, a large portion of generated materials may be charge unbalanced, and existing methods either fix the composition in advance or discard charge unbalanced samples after generation. In this work, we propose AMGenC, a new generative inverse design method for amorphous materials that can guarantee the generation of formally charge balanced samples, with minimal additional computational overhead and without sacrificing inverse design accuracy. AMGenC achieves this through an element noise that gives the generation process a starting point centered around charge balance, and the combination of a per-step soft projection and a final discrete projection for steering the elements toward exact charge balance throughout the generation. We perform extensive experiments on two amorphous materials datasets. Experimental results provide evidence that AMGenC achieves its design goal.
Dialectal Arabic speech recognition remains constrained by the scarcity of dialect-resolved transcribed speech despite the abundance of web audio. We introduce GigaSpeech 3, a corpus of 39,000 hours and 10 million weakly transcribed speech segments covering Algerian, Egyptian, Iraqi, Saudi, and Syrian Arabic. Starting from 60 hours of manual transcription per dialect, we train labeling models using speech-recognition-biased self-supervised learning and apply them to dialect-matched web audio. Data selection and four rounds of noisy-student refinement for Saudi and Syrian Arabic yield the final corpus. We validate its utility through dialect-specific and multi-dialect speech recognition training and cross-domain evaluation. On GigaSpeechBench, the final dialect-specific models outperform their labeling models across all five dialects and achieve a macro-average word error rate of 23.86%, a 16.6% relative reduction from the strongest third-party system covering all five dialects, while outperforming the best third-party result on three dialects. A multi-dialect model using self-predicted dialect identities achieves a macro-average word error rate of 25.51% while also supporting Modern Standard Arabic. Across eight conditions from three external corpora, a GigaSpeech 3-trained Zipformer achieves the lowest word error rate among six evaluated models in every condition, while adapting the Qwen3 speech recognizer on GigaSpeech 3 improves performance across all eight conditions. We release the corpus and recipes as a foundation for dialect-aware Arabic speech processing.
Video Large Language Models (Video-LLMs) have achieved strong performance on many video understanding tasks, yet remain vulnerable to temporal hallucination in time-sensitive reasoning scenarios such as scene transition understanding and event ordering. Existing solutions either require costly retraining or introduce inefficient multi-pass inference and external supervision at test time. In this work, we show that temporal evidence is already encoded in Video-LLM hidden representations but is not fully exploited during standard inference. Motivated by this observation, we propose Single-Pass Orthogonal Decomposition (SPOD), a training-free framework that recovers suppressed temporal evidence through analytical hidden-state manipulation in a single forward pass. SPOD consists of two components: Frame Distinctiveness Analysis (FDA), which identifies salient frames from middle-layer hidden-state geometry, and Temporal Delta Subspace (TDS), which constructs orthogonal temporal and prior subspaces from weighted inter-frame hidden-state differences. By decomposing final hidden states into temporal, prior, and residual components and rescaling them independently, SPOD enhances dynamic information, suppresses static appearance bias, and preserves complementary information. Experiments on VidHalluc, EventHallusion, and VideoMMMU show that SPOD consistently outperforms strong inference-time baselines across temporal reasoning tasks and model architectures while retaining near-vanilla inference efficiency. These results suggest that improving temporal reasoning in Video-LLMs may depend not only on additional supervision, but also on recovering temporal evidence already present in their internal representations. Our code is available at https://github.com/VioAgnes/SPOD.
Tim Knappe, Ryan Li, Ayush Chauhan, Kaylee Chhua, Kevin Zhu, Sean O'Brien
While large language models (LLMs) have rapidly improved their performance on a broad number of tasks, they still often fall short on reasoning tasks. As LLMs become more integrated in diverse real-world tasks, advancing their reasoning capabilities is crucial to their effectiveness in nuanced, complex problems. Wang et al.'s self-consistency framework reveals that sampling multiple rationales before taking a majority vote reliably improves model performance across various closed-answer reasoning tasks. Standard methods based on this framework aggregate the final decisions of these rationales but fail to utilize the semantic information detailed in the step-by-step reasoning paths. Our work introduces semantic self-consistency, enhancing this approach by incorporating and analyzing both the reasoning paths of these rationales in addition to their final decisions before taking a majority vote. These methods not only improve the reliability of reasoning paths but also cause more robust performance on complex reasoning tasks.
Andreas Bjerregaard, Rachael DeVries, Dionysia Danai Brilli, Panagiotis Antoniadis, Yan Li, Anders Krogh, Wouter Boomsma, Ole Winther 等 9 位作者
Genomic foundation models are increasingly used to propose DNA sequences and edits. But successful design requires capabilities that current benchmarks do not measure. Here we introduce BEND 2.0 (BENchmark for DNA language models), which tests how well models predict the effects of DNA changes and whether these predictions can guide sequence design. Across eight tasks built on five experimental datasets, we find that model choice and how representations are accessed both affect performance. Linear readouts fitted to frozen representations can match or exceed models' native outputs, and functional post-training substantially improves the models' capabilities. Linear readouts transfer well across experimental assays and genes, suggesting that models capture functional patterns beyond the data used to fit individual predictors. At the same time, strong prediction of average effects can conceal weaker prediction of context-specific effects, and accurate prediction of combined mutation effects does not imply accurate prediction of their interactions. In design tasks, model predictions can reduce the number of candidates needed to find a desired response, but still leave substantial room for better design choices, particularly for context-specific and multi-cell objectives. Our benchmark reveals that genomic models now display signs of general utility but have yet to saturate on performance in even simple tasks.
Sampling and selection spend test-time compute inside a candidate pool, but they cannot recover a correct answer that never enters it. We introduce ValueFormation, a test-time policy that decides after each primary-model draw whether to stop, continue sampling, or acquire one proposal from a heterogeneous reasoning specialist. The three actions are compared on one cost-regularized value scale learned from vote-state features; the calibration is frozen across tasks and primary models, and acquired proposals are accepted only through a frozen posterior comparison. On held-out MATH-500, one Llama-fitted calibration improves over the strongest of eight identically evaluated baselines by more than fifteen points for each of three 8B primaries and transfers unchanged to Qwen3-14B. Removing acquisition lowers Llama MATH accuracy by 20.43 points, while same-trace and same-access replays attribute the gain to state-conditioned acquisition rather than specialist access alone. Where primary coverage is high, as on GSM8K, acquisition is rare and gains mainly come from cheaper allocation; on MBPP, a fixed code route attains the top accuracy at the lowest interaction cost. Varying the token price monotonically reduces calls without retraining.
Bingsen Chen, Boyan Li, Ping Nie, Yuyu Zhang, Xi Ye, Chen Zhao
Existing benchmarks for Deep Research Agents (DRAs) treat report generation as a single-shot writing task, which fundamentally diverges from how human researchers iteratively draft and revise reports via self-reflection or peer feedback. Whether DRAs can reliably revise reports with user feedback remains unexplored. We introduce Mr Dre, an evaluation suite that establishes multi-turn report revision as a new evaluation axis for DRAs. Mr Dre consists of (1) a unified long-form report evaluation protocol spanning comprehensiveness, factuality, and presentation, and (2) a human-verified feedback simulation pipeline for multi-turn revision. Our analysis of five diverse DRAs reveals a critical limitation: while agents can address most user feedback, they also regress on 16-28% of previously covered content and citation quality. Over multiple revision turns, even the best-performing agents leave significant headroom, as they continue to disrupt content outside the feedback's scope and fail to preserve earlier edits. We further show that these issues are not easily resolvable through inference-time fixes such as prompt engineering and a dedicated subagent for report revision.
On-device multimodal agents are limited by two deployment mismatches: pretrained attention structures may not divide evenly across the compute cores of an edge accelerator, causing irregular partitions; and standard prefix caching keeps historical visual tokens in the persistent KV cache, whose capacity grows with the number of rounds. We present PARC, which combines parallelism-aligned structured pruning with lifetime-aware KV-cache reorganization. At the model level, it uses first-order Taylor importance and grouped-query-attention-aware selection to make attention dimensions divisible by the target core count while preserving the query-to-key/value ratio and head dimension. At runtime, it separates the system prompt, persistent text history, and transient visual observations by lifetime, updating text history during tool execution so that historical images leave the persistent cache and more interaction rounds fit a fixed budget. On 2,062 GUI actions, the incremental interface shows no accuracy regression, and task-aware recovery training restores 98.7% of the action-type accuracy gap and the coordinate hit rate to the original level. Under seven-core deployment, hardware-aligned pruning reduces vision-encoder latency by 11.4%, improves decoding throughput by 23.8%, and lowers per-heap high-water memory by 12.7%; a parameter-matched control confirms that the speedup stems mainly from core alignment rather than parameter reduction.
Jing Zhang, Ganxuan Yang, Yifei Yang, Siqi Wen, Baiyan Li, Zhengyang Qiu, Hasan Ahmed, Muhammad Bilal
Searchable Encryption (SE) enables users to query outsourced encrypted data while preserving data confidentiality. However, most efficient schemes still leak the search pattern and access pattern, which may allow an honest-but-curious cloud server to infer query contents, user interests, or returned records from repeated searches and observed results. Existing pattern-hiding solutions mainly target keyword queries and do not naturally support Boolean range queries over encrypted spatial data. This paper presents BRASP, a searchable encryption scheme for Boolean range queries over encrypted spatial data. BRASP combines Hilbert-curve-based prefix encoding with encrypted prefix--ID and keyword--ID inverted indexes to support efficient spatial range filtering and conjunctive keyword matching. To hide the search pattern and access pattern under a dual-server setting, BRASP integrates index shuffling for encrypted keyword and prefix entries with ID-field redistribution across two non-colluding cloud servers. BRASP also supports dynamic updates and achieves forward security. We formalize the security of BRASP through confidentiality, shuffle indistinguishability, query unforgeability, and forward-security analyses, and we evaluate its performance experimentally on a real-world dataset. The results show that BRASP effectively protects query privacy while incurring relatively low computation and communication overhead. To facilitate reproducibility and further research, the source code of BRASP is publicly available at https://github.com/Egbert-Lannister/BRASP
Junyan Liu, Haipeng Luo, Lillian J. Ratliff, Zihan Zhang
We study online learning in two-player uninformed Markov games, where the opponent may follow an arbitrary, unknown policy in each episode and the learner does not observe the opponent's actions. [TWYS21] showed that achieving sublinear external regret requires an exponential dependence on the episode length $H$. Motivated by this limitation, they studied the weaker notion of Nash-value regret and obtained an $\widetilde{O}(K^{2/3})$ regret bound, where $K$ is the number of episodes. More recently, [LLZR26] introduced a stronger notion called empirical Nash-value regret and achieved an $\widetilde{O}(\sqrt{K})$ bound when the opponent's non-stationarity is moderate, including the case where the opponent follows a fixed policy. However, whether the $\sqrt{K}$ rate is achievable in the worst case remained open. In this paper, we resolve this open question by developing an algorithm that achieves $\widetilde{O}(H^{3/2}S\sqrt{AK})$ empirical Nash-value regret against arbitrary opponent policies, where $S$ is the number of states and $A$ is the number of learner's actions. Specifically, our algorithm combines global and local planning by maintaining a collection of candidate predictions of future losses, each associated with a policy. Global planning learns which candidates to favor across episodes, while local planning combines their policies differently at each state to make decisions. A key ingredient is to coordinate these two components to control the discrepancy between local loss estimates and the total loss of the executed policy, using only the learner's observed trajectories.
Yaxin Gong, Chongming Gao, Gangyi Zhang, Chenxiao Fan, Leyang Shen, Jiakai Wang, Dong Wang, Yang Liu 等 10 位作者
Multi-step Web tasks often require several LLM agents to gather and combine evidence. The communication graph of a fixed workflow restricts which agents can read each other's outputs: evidence retrieved by one agent may never reach the agent that needs it, leaving an information gap and an incomplete or incorrect answer. Adding a connection can deliver the missing evidence, but can also lead the receiver to adopt an incorrect upstream conclusion that displaces a correct answer; the final output alone does not separate the two. We propose Selective Probing and Link Insertion for Closing information gaps in multi-agent Execution (SPLICE), which locates information gaps per query and decides what to adopt after adding an edge. After the original workflow runs, SPLICE uses the query and node outputs to locate a possible gap, adds one directed edge from a source to a receiver to deliver the relevant evidence, and re-executes the receiver and affected downstream nodes, producing a candidate answer. The adoption decision reads both the final answers of the two runs and the process changes the edit caused inside the workflow: how far the receiver moved, where its new content came from, and whether the change reached the final answer. A selector trained on these process features chooses between adopting the candidate answer and retaining the original. On five benchmarks, SPLICE raises the task score on Web tool use by 8.0% and the mean score on four reasoning and planning benchmarks by 5.7% relative to the static skeleton, at 1.25 to 1.44$\times$ the skeleton's cost, and transfers zero-shot to four unseen backbones.
Chenxiao Fan, Chongming Gao, Gangyi Zhang, Yaxin Gong, Leyang Shen, Jiakai Wang, Dong Wang, Yang Liu 等 12 位作者
Complex information seeking on the Web often requires collecting evidence from multiple documents and reasoning across them, as in multi-hop question answering. Multi-agent systems address such tasks by dividing work among specialized agents, where upstream agents produce intermediate answers and supporting evidence that downstream agents build upon. However, upstream answers vary in correctness, and most systems provide no quantitative signal about it, leaving the downstream agent to judge whether to trust or verify what it receives based solely on the textual evidence. In this paper, we propose MASCAL. Each agent learns to produce a correctness probability for its answer by jointly optimizing task performance and confidence calibration in reinforcement learning; the upstream agent passes this probability at handoff, giving the downstream agent a quantitative basis for its decisions, and the final answer also carries a calibrated probability, quantifying the reliability of the system output. To guide the downstream agent in using this signal during training, we introduce confidence-aware weighting, which increases the learning weight at positions responsive to the upstream probability, so that the agent learns more from decisions where it actively uses the signal. Across three collaboration architectures spanning Web retrieval and reasoning tasks, MASCAL improves task accuracy and calibration quality for both agents, and generalizes to deeper collaboration chains and different model scales. Intervention analysis confirms that upstream confidence affects downstream decision behavior, and the learned final confidence supports selective answering and targeted rechecking.
Nuclei instance segmentation is fundamental to quantitative pathology. Yet high-accuracy methods mostly build on large vision transformers (ViTs) or heavy convolutional backbones, whose parameter count and compute cost hinder deployment in clinical and edge settings. We present NucShare, a lightweight model whose shared decoder handles nuclear pixel segmentation, horizontal-vertical distance, and classification through learnable transposed convolutions for upsampling and three lightweight heads, using 37% of independent decoders' parameters with no loss in accuracy. Output-level distillation then transfers knowledge from NucShare-T (UNI2-h) to NucShare-S (ConvNeXt V2-Tiny), reducing 690.5M parameters to 34.7M. On PanNuke they reach mPQ 0.524 and 0.519, above the prior best CFR-SAM-H at 0.511, and NucShare-S surpasses HoVer-Net and Mask R-CNN by 5.6 and 15.0 points; on zero-shot MoNuSeg it attains bPQ 0.668 against CFR-SAM-H's 0.662. Code: https://github.com/L-iMeng-L/Newidea
Huanyue Liao, Seok Min Lim, Yan Ling Lim, Shi Min Cheng, Kellin Pelrine, Chen Hui Ong
AI companion chatbots are increasingly used by vulnerable individuals, including adolescents in suicidal crisis, for emotional support. Existing safety evaluations focus on base models, leaving gaps in product-level assessment, crisis-stage sensitivity, and comparisons between companion apps and general-purpose LLMs. We introduce MIRROR-Eval (Multi-turn Interaction Risk Review: Observing suicidal-crisis Responses), a dataset of 960 annotated 20-turn conversations across eight deployed platform configurations, six clinically grounded personas, and two crisis stages: ideation and explicit planning. An LLM-as-a-Judge pipeline applies an expert-reviewed rubric covering user risk severity, chatbot risk awareness, and intervention behavior. We observe a three-tier harm structure: companion apps (Character.AI and Replika) and Grok 4.3 form a high-harm tier; Gemini 3.1 Pro is mid-range; GPT-5.5 and Claude Sonnet 5 show substantially lower harm. At planning, companion apps show increased glorification and normalisation, advancement toward harmful action, and harmful methods and instructions. Crisis recognition and empathetic language do not ensure safe responses. These findings highlight the need for product-level safety evaluation across crisis stages. Dataset and code: https://huggingface.co/datasets/imda-biztech/MIRROR-Eval. Warning: this paper includes examples that may be offensive or harmful.
Renzhi Wu, Li Sheng, Zikun Cui, Junjie Yang, Tai Guo, Keke Zhai, Yan Li, Jerry Fu 等 11 位作者
Long-sequence recommendation can improve along two axes: \emph{evidence scaling} exposes longer histories, while \emph{reasoning scaling} adds layers that interpret that evidence for a candidate. Existing models couple them: sequence encoders propagate every added event through depth, whereas one-shot selection prevents later layers from revisiting omitted evidence. We introduce the Latent Workspace Ranker (LWR), which keeps typed evidence in persistent banks and propagates a fixed-size candidate-local workspace. Each layer rereads the banks from its evolving workspace and archives compact post-interaction state, allowing history length and reasoning depth to grow independently. We compare four architectures across various history--depth configurations using evidence scaling at fixed depth, reasoning scaling at fixed 10K history, and their joint compute-efficient frontier. LWR leads all families at long histories, continues improving with reasoning compute, and converts additional training compute into quality gains at the fastest fitted rate on the joint frontier. At 10K history, it improves NE by 0.19% over HSTU while using 69.6% fewer forward GFLOP/request. It also outperforms existing baselines on two billion-scale ranking surfaces. These results show that separating addressable evidence from propagated state makes reading more and reasoning deeper independently controllable and jointly useful.
Machine unlearning aims to reproduce the behavior of a model trained without designated data. For probabilistic decision models, this target also depends on the candidate options supplied at inference. We introduce DecisionUnlearn-Bench, an empirical framework that compares full, unlearned, and retain-only models across deletion granularities, label renderings, and candidate sets. Across 20 eligible model configurations on the BANKING77 dataset, full-to-oracle displacement increases consistently from example to association to class deletion. Replacing semantic option labels with opaque codes nearly eliminates the retrained oracle's recognition of deleted classes, even with successful acquisition of the full model, whereas changing the candidate set alters both the oracle and the measured unlearning progress. SimNPO satisfies the screening criterion in 19 of 20 configurations, but its gains vary substantially across seeds and deletion targets. Development tuning on the CLINC150 dataset does not establish a positive pooled effect across 48 intents. At approximately matched suppression, methods can differ markedly in oracle fidelity. Recovery experiments further distinguish initial forgetting from subsequent adaptation and show conclusions that depend on both architecture and metric. These results show that evaluating decision-model unlearning relative to a schema-specific retraining target, with separate reporting of effect size, utility, robustness, and recoverability.
Skills, as reusable external knowledge, have demonstrated substantial potential to enhance the ability of Large Language Models (LLMs) to solve complex tasks. However, existing skill-augmented generation methods typically rely on separate models for skill retrieval and skill utilization, resulting in disconnected optimization of these two processes. To address this gap, we propose Skill-Uni, a unified reinforcement learning framework that jointly optimizes skill retrieval and utilization. Specifically, we adopt a generative skill retrieval paradigm that enables a single model to both learn how to retrieve relevant skills and leverage them for task solving. To effectively train this unified framework, we introduce a two-stage training strategy consisting of warm-up instruction tuning and GRPO, allowing the model to progressively improve its skill retrieval and utilization capabilities. Extensive experiments on five skill-related benchmarks demonstrate the effectiveness of Skill-Uni, yielding average performance gains of 66.59% and 34.82% over the base Qwen3-4B and LLaMA3.1-8B models, respectively. Besides, Skill-Uni is compatible with various skill retrieval engines, demonstrating its flexibility and generalizability. Additional in-depth analyses also provide insights into the effectiveness of Skill-Uni.
Fang Sun, Hekai Cheng, Ruoyan Li, Yuanzhou Chen, Wei Wang, Jason Cong, Yizhou Sun
Kohn-Sham density functional theory (DFT) is the standard method for computing the electronic structure of molecules and materials, but its self-consistent-field (SCF) iterations make it expensive for large systems and high-throughput screening. Machine-learned Hamiltonian models offer a shortcut: an equivariant network predicts the Kohn-Sham Hamiltonian matrix in one forward pass, and diagonalizing the prediction yields orbital energies, the HOMO-LUMO gap, molecular orbitals and the electron density. These models are trained to match the reference matrix entry by entry, yet a small entry-wise error does not guarantee accurate eigenvalues or eigenvectors. Two physics-informed alternatives go partway: a recently proposed wavefunction-alignment loss adds a spectral penalty without diagonalizing the prediction, and matching the electron density of the diagonalized prediction supervises it directly. Neither is complete: by representing the prediction error in the eigenbasis of the reference Hamiltonian, we show that each sees only part of the error, and that shifting the predicted orbital energies leaves the electron density exactly unchanged. The same analysis identifies a complete training signal: the cumulative local density of states (cumulative LDOS), the number of electrons at each point in space with energy below a given threshold. The cumulative LDOS is a single physical observable that contains the electron density as a special case, responds to every component of the error, determines the Hamiltonian uniquely, and yields gradients that do not vanish for large orbitalenergy errors. Fine-tuning three Hamiltonian networks on QH9 in this way, which we call FermiScope, reduces the HOMO-LUMO gap error without increasing the element-wise error. The gains carry over to out-of-distribution and molecular-dynamics splits, where the wavefunction-alignment loss helps less.
Mengyao Lyu, Yan Li, Huasong Zhong, Hui Chen, Jungong Han, Guiguang Ding, Zhenheng Yang
Recent research demonstrates that pretrained large language models (LLMs) require minimal but carefully curated data for instruction-tuning. Building on LLMs, multi-modal models further contend with far larger data volumes, heterogeneous sources, and cross-modal alignment, amplifying both the necessity and the complexity of data selection. However, existing methods rely on deeply extracted embeddings, gradients or narrow quality scores that are time-consuming to compute and limited in expressiveness, frequently failing to surpass a simple random-sampling baseline. To overcome these limitations, we step back from complex embedding spaces and target the fundamental objective of instruction tuning: how and what knowledge should be elicited from the pretrained model and exposed to users. We leverage interaction \underline{S}tyles as a proxy for structural diversity, and decompose data quality into capability-specific \underline{S}cores, yielding an explicit yet \underline{R}ich (SSR) profile for each sample. We further cast the selection as submodular coverage maximization over the capability--style grid, providing an approximation guarantee. SSR consistently delivers superior performance across two distinct architecture stacks, LLaVA-1.5 (Vicuna + CLIP) and LLaVA-OneVision (Qwen2 + SigLIP), two large-scale data pools, six data budget settings, and 14 multi-modal benchmarks. To our knowledge, SSR is the first data selection method validated at million-scale, achieving 99.1% of full-data performance using only 30% of a 2.6M pool. Furthermore, the SSR profile models and selected subsets transfer training-free across data pools and model architectures with capability extensibility.
We present CPG-ENN, a trainable quadruped locomotion framework that combines explicit neural coordination with learned motion realization. A symmetry-based eight-neuron central pattern generator (CPG) supplies gait rhythms, while a pretrained, leg-shared execution neural network (ENN) maps rhythmic signals and trajectory parameters to joint-angle references without online inverse kinematics. A single feedback policy modulates six physically interpretable trajectory variables. The controller tracks planar velocity and yaw-rate commands across walk, trot, pace, bound, and pronk without gait-specific learned controllers or transitionspecific rewards. Simulation yields 99.39% transition success across all 20 directed gait pairs and all source-target speed combinations from {0.5, 1.0, 1.5, 2.0} m/s. Ablations show high cross-gait transition feasibility without transition exposure; exposure mainly refines gait acquisition and velocity transients. Comparisons with AllGaits and Walk These Ways demonstrate accurate gait-speed regulation, and Unitree Go2 experiments validate five gaits and their transitions.
Vision-Language-Action (VLA) models combine multimodal perception with robotic control, demonstrating remarkable potential for generalizable manipulation. Yet, enhancing their execution typically relies on auxiliary sensory inputs (e.g., depth, 3D point clouds) or costly fine-tuning. Inspired by the insight that Feed-Forward Networks (FFNs) act as key-value memory, we propose Action-Entropy-Weighted and Attention-Guided Retrieval of Observation Evidence (AWARE), an effective, training-free, and plug-and-play module for VLAs. Specifically, AWARE tracks layer-wise uncertainty via an element-wise Action Entropy metric, and retrieves observation evidence into the element's FFN through associative memory lookup. Crucially, the retrieved features are weighted by instruction-guided cross-modal attention to isolate task-relevant observation cues, while injection strength is adaptively scaled by the element's entropy and relative network depth. This selectively channels observation evidence to high-entropy action dimensions and deep layers where grounding decays, while leaving converged representations intact. Across LIBERO, SIMPLER, CALVIN, and real-world robot manipulation, AWARE consistently improves diverse VLA architectures with minimal computational overhead, achieving the largest gains on long-horizon tasks without architectural changes or additional training. Project page: https://jiabingyang01.github.io/AWARE/.
Wei-Wei Zhang, Pengkai Wang, Yan Li, Zhijie Sang, Guan-Qun Zhou, Jun-Yan Li, Shu-Han Zhou, Pei-Wen Liu 等 16 位作者
Large language models (LLMs) show promise in specialty medicine, yet existing evaluations largely focus on isolated clinical encounters rather than tracking how clinical responsibilities evolve across a longitudinal care pathway. Here we introduce NPC-RubricBench, a phase-aware benchmark for evaluating multistage clinical reasoning across the cancer care continuum, using nasopharyngeal carcinoma as a defined disease context. Curated from longitudinal records of 405 patients across two cancer centers, the benchmark comprises 941 patient-phase units across six sequential care phases and 17,688 patient-specific atomic criteria spanning five evaluation dimensions, with each assessment strictly restricted to decision-time clinical evidence. Evaluating 30 general-purpose and medical LLMs revealed that strong diagnostic performance does not carry over to downstream care. We evaluated 30 general-purpose and medical LLMs, showing that high diagnostic performance did not translate into improved downstream care. Instead, the dominant failure modes reversed across phases. Explicit decision errors dominated anatomical staging at 48.2%, whereas decision omissions dominated initial work-up and post-treatment follow-up at 33.4% and 30.4%, respectively. In the evaluated model set, medically adapted models showed higher omission rates than general-purpose models. Moreover, varying query granularity exposes a critical specification trade-off: broad open-ended requests degrade Decision Quality across all 30 models, while more detailed, checklist-like prompts raise Decision Quality but reduce Evidence Faithfulness and Safety Adherence. Furthermore, NPC-RubricBench demonstrates that static diagnostic benchmarks are insufficient to ensure consistent clinical reliability over time, highlighting the importance of phase-adaptive safeguards in multistage medical AI.
Generating humorous memes is a challenging multimodal task that moves beyond direct image-to-caption supervision. It requires a nuanced reasoning over visual content, contextual cues, and subjective humor. To bridge this gap between visual perception and humorous punchline creation, we propose HUMOR}, a novel framework that guides VLMs through hierarchical reasoning and aligns them with group-wise human preferences. First, HUMOR employs a hierarchical, multi-path Chain-of-Thought (CoT): the model begins by identifying a template-level intent, then explores diverse reasoning paths under different contexts, and finally anchors onto a high-quality, context-specific path. This CoT supervision, which traces back from ground-truth captions, enhances reasoning diversity. We further analyze that this multi-path exploration with anchoring maintains a high expected humor quality, under the practical condition that high-quality paths retain significant probability mass. Second, to capture subjective humor, we train a pairwise reward model that operates within groups of memes sharing the same template. Following established theory, this approach ensures a consistent and robust proxy for human preference, even with subjective and noisy labels. The reward model then enables a group-wise reinforcement learning optimization, guaranteeing providing a theoretical guarantee for monotonic improvement within the trust region. Extensive experiments show that HUMOR empowers various VLMs with superior reasoning diversity, more reliable preference alignment, and higher overall meme quality. Beyond memes, our work presents a general training paradigm for open-ended, human-aligned multimodal generation, where success is guided by comparative judgment within coherent output group.
Jiaming Zhang, Fuyao Zhang, Pengjun Xie, Wei Yang Bryan Lim
Deep neural networks are known to rely on non-robust features that align poorly with human perception. Prior work has largely framed these features as opportunistic shortcuts that help standard models on clean data. We show instead that non-robust cues can be antagonistic: they can act as Natural Jammers that actively mislead standard models on clean, unmodified images. Using disagreement between a standard model and robust models as an analysis lens, we identify a Jammed Set on which robust models outperform standard models. We provide causal evidence for Spectral-spatial Locking: jamming arises only when high-frequency components are precisely aligned with specific spatial structures. Disrupting this lock through spectral filtering or small geometric shifts can restore correct predictions. Crucially, we uncover a Substitution Paradox: while individual jammers are fragile to geometric shifts, the aggregate prevalence of jamming remains approximately stable and is redistributed rather than eliminated under such transformations. Our results reframe natural non-robustness as a failure field over the empirical model-data geometry: one that redistributes rather than disappears under small transformations.
Ge Chang, Jinbo Su Renmin University of China, Jiacheng Liu, Pengfei Yang, Yuhao Shang, Huiwen Zheng, Hongli Ma GDS, Yan Liang 等 10 位作者
Integrating textual graphs into Large Language Models (LLMs) is promising for complex graph-based QA. However, a key bottleneck is retrieving informative yet compact subgraphs that fit the LLM context. Existing retrievers often struggle, relying either on shallow embedding similarity or costly interactive policies that require excessive supervision.To address these challenges, we introduce Graph-S^3, an agentic textual graph reasoning framework featuring an LLM-based retriever trained with synthetic stepwise supervision. Rather than relying on final answer rewards—which often yield sparse and unstable signals—we optimize the retriever by evaluating each step against offline-extracted golden subgraphs.Our approach distills golden subgraphs via a specialized data synthesis pipeline to formulate dense rewards, facilitating a two-stage training scheme that effectively learns the interactive graph exploration policy.Based on extensive experiments on three common datasets in comparison with seven strong baselines, our approach achieves an average improvement of 15.6% in accuracy and 17.2% in F_1 score. The advantage is even higher in more complicated multi-hop reasoning tasks.
尚无 AI 解读，展示原文摘要
cs.CL · cs.AI
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yan Liu, Feng Zhang Peking University, Zhanyu Ma, Jun Xu Meituan, Jiuchong Gao, Jinghua Hao, Renqing He, Han Liu Dalian University of Technology 等 9 位作者
High-quality chain-of-thought has demonstrated strong potential for unlocking the reasoning capabilities of large language models. However, current paradigms typically treat the reasoning process as an indivisible sequence, lacking an intrinsic mechanism to quantify step-wise information gain. This granularity gap manifests in two limitations: inference inefficiency from redundant exploration without explicit guidance, and optimization difficulty due to sparse outcome supervision or costly external verifiers. In this work, we propose CoT-Flow, a framework that reconceptualizes discrete reasoning steps as a continuous probabilistic flow, quantifying the contribution of each step toward the ground-truth answer. Built on this formulation, CoT-Flow enables two complementary methodologies: flow-guided decoding, which employs a greedy flow-based decoding strategy to extract information-efficient reasoning paths, and flow-based reinforcement learning, which constructs a verifier-free dense reward function. Experiments on challenging benchmarks demonstrate that CoT-Flow achieves a superior balance between inference efficiency and reasoning performance.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yelin Chen Xinjiang University, Fanjin Zhang Renmin University of China, Suping Sun, Yunhe Pang, Yuanchun Wang, Jian Song, XiaoYan Li, Lei Hou Tsinghua University 等 11 位作者
Understanding research papers remains challenging for foundation models due to specialized scientific discourse and complex figures and tables, yet existing benchmarks offer limited fine-grained evaluation at scale. To address this gap, we introduce RPC-Bench, a large-scale question-answering benchmark built from review–rebuttal exchanges of high-quality computer science papers, containing 15K human-verified QA pairs. We design a fine-grained taxonomy aligned with the scientific research flow to assess models’ ability to understand and answer why, what, and how questions in scholarly contexts. We also define an elaborate LLM–human interaction annotation framework to support large-scale labeling and quality control. Following the LLM-as-a-Judge paradigm, we develop a scalable framework that evaluates models on correctness-completeness and conciseness, with high agreement to human judgment. Experiments reveal that even the strongest models (GPT-5) achieve only 68.2% correctness-completeness, dropping to 37.46% after conciseness adjustment, highlighting substantial gaps in precise academic paper understanding.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Human infants, with only a few hundred hours of speech exposure, acquire basic units of new languages, highlighting a striking efficiency gap compared to the data-hungry self-supervised speech models. To address this gap, this paper introduces SpidR-Adapt for rapid adaptation of speech units to new languages using minimal unlabeled data. We cast such low-resource speech representation learning as a meta-learning problem and construct a multi-task adaptive pre-training (MAdaPT) protocol which formulates the adaptation process as a bi-level optimization framework. To enable scalable meta-training under this framework, we propose a novel heuristic solution, first-order bi-level optimization (FOBLO), avoiding heavy computation costs. Finally, we stabilize meta-training by using a robust initialization through interleaved supervision which alternates self-supervised and supervised objectives. Empirically, SpidR-Adapt achieves rapid gains in phonemic discriminability (ABX) and downstream spoken language modeling scores (sWUGGY, sBLIMP, tSC), surpassing in-domain toplines after training on less than 1h of target-language audio and delivering 100\times greater data efficiency than standard multi-task training.. These findings highlight a practical, architecture-agnostic path toward biologically inspired, data-efficient representations. We open-source the training code and model checkpoints at https://github.com/facebookresearch/spidr-adapt.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表