Causal-prior-constrained deep learning for rapid-intensification warning of western North Pacific tropical cyclones
Xinrui Li, Yijun Jin, Yifan Dong, Shoujuan Shu
AI 速览：作者提出CausalVisionRI，将观测发现的因果图作为可审计先验约束融入深度学习，用于西北太平洋热带气旋24小时快速增强预警。
Rapid intensification (RI) is one of the most dangerous and unpredictable phases of tropical cyclones, and its environmental and structural drivers are difficult to identify and separate in observations. Here we introduce CausalVisionRI, a causal-prior-constrained deep-learning system for 24-h RI warning in the western North Pacific. It integrates infrared remote sensing imagery, atmospheric and oceanic fields, and an observationally discovered causal graph into a unified model. The graph and eight sign-verified effects on 24-h intensity change provide an auditable causal prior for training. Discovered from 522 cyclones in 2000-2017, it identifies causal pathways linking atmospheric and oceanic conditions, the cyclone's dynamical and structural state, and its 24-h intensity change. On a forward holdout covering 2020-2024 and 3,004 records, the five-member ensemble achieves a Brier skill score (BSS) of 0.259 relative to a fixed development climatology, significantly exceeding four reference systems; the three LDA baselines reach at most 0.105. It also attains higher detection rates than a published knowledge-block scheme and a reconstructed contrastive-learning scheme at matched false-detection rates under our USA 30-kt standard and the published RITCF standard. Ablations show that the graph pathway adds 0.032 BSS across the holdout and 0.063 over 2020-2021, with a significant contribution in one ensemble member. A deep-learning model constrained by observationally discovered causal structure thus offers a new route to physically informed RI warning.
Recent advances in autoregressive video diffusion have enabled sequential and streaming video generation. However, long-horizon generation requires increasingly large KV caches, making efficient compression without sacrificing quality challenging. Existing methods mostly select historical frames based on attention scores, but their context decisions remain coarse. When multiple frames are generated in the same chunk, these methods often apply a shared history selection to the whole chunk, score historical frames solely by attention, and assign head-wise budgets either uniformly or by attention-pattern heuristics rather than explicit head-importance estimation. We show that frames within the same generated chunk can depend on distinct historical frames, that the same historical frame can receive different attention scores as its relative temporal distance to the current frames changes, and that masking different heads induces unequal generation degradation. Motivated by these findings, we propose Focused Forcing, a training-free KV selection method that focuses cached history along both generated-frame and head dimensions. For each generated frame, Focused Forcing preserves the most relevant and distinctive historical frames by combining attention scores with diversity scores of historical frames, while assigning larger budgets to heads with higher estimated importance. Across multiple autoregressive generation paradigms, Focused Forcing achieves up to 1.48x end-to-end acceleration without training, while improving visual quality and text alignment. Our code is available at https://github.com/caipeiliang2004/Focused-Forcing.
Qin-Wen Luo, Sheng Ren, Xiang Chen, Rui Liu, Jun Fang, Naiqiang Tan, Jun Wang, Hua Chai 等 9 位作者
Deep search agents extend language models by interleaving reasoning with external retrieval. However, effective search-agent training requires addressing two challenges beyond final-answer optimization. First, the capability frontier of agents is largely determined by training questions, whereas fixed or labeled task collections cannot adapt as agents improve. Second, outcome-only reinforcement learning treats all successful trajectories equally, even when some correct answers rely on redundant searches, creating an inherent tension between answer accuracy and tool efficiency. To address these issues, we introduce \textsc{ZEST-Search}, a zero-label and tool-efficient framework for search-agent training via heterogeneous co-evolution. Starting from unlabeled reference questions, \textsc{ZEST-Search} constructs validated search tasks online through an answer-first and evidence-path-guided generation process. A heterogeneous proposer adaptively constructs training tasks around the solver's evolving capability boundary, while the solver learns search-augmented reasoning through reinforcement learning. To jointly optimize task difficulty and search efficiency, \textsc{ZEST-Search} employs an intermediate-difficulty reward for the proposer and a task-local tool-aware reward for the solver, encouraging informative tasks and correct answers with fewer redundant searches. Experiments on seven general-domain and multi-hop question answering benchmarks demonstrate that \textsc{ZEST-Search} consistently improves search-agent performance while achieving higher tool efficiency without labeled search-training data or expert search trajectories. Further analyses show that heterogeneous generation, adaptive difficulty control, and tool-aware optimization enable more effective search-agent training.
Khai Le-Duc, Nguyen Kim Hai Bui, An Ngo, Nam D. Hoang, Yueqi Song, Jiarui Liu, Thai Quoc Hoang, Fang Wu 等 14 位作者
Vision-language models can get individual judgments right while struggling to recover and consistently use the exact properties of a 3D object. To characterize this gap, we introduce BedBench, a controlled benchmark of thirteen procedurally generated tasks spanning individual judgments and composite answers. Across twelve open-weight and four frontier models, per-color counting exposes a pronounced weakness: humans achieve 93.3% accuracy, compared with 7.2% for the best model. Targeted fine-tuning raises one model's accuracy from 3.0% to 68.6%, but reveals a further difficulty: its answers about the most common colors remain inconsistent with its own per-color counts. Supplying those counts as text improves its judgments, showing that recovering a property does not mean using it consistently. Reasoning, scaling, and added visual cues each improve individual judgments, but none makes exact answers reliable. To locate the difficulty, we probe Qwen3.5-27B: exact per-color counts remain recoverable at the language model's input, yet the model's answers rarely recover them (65.0% for the probe vs. 3.0% for the model). Image-removal controls further show that some apparent successes require no visual evidence. These findings motivate evaluating 3D reasoning by exact correctness, consistent use, and image dependence, rather than accuracy alone. Code and dataset are publicly available: https://github.com/leduckhai/BedBench
Proactive agents must decide whether, when, and how far to assist without an explicit request. Existing evaluations often rely on automated scores or binary verdicts, offering limited human diagnosis of evidence use and intervention quality. To address these gaps, we introduce HALO, Human-Grounded GrAnular EvaLuation of PrOactive Agents. HALO-Arena covers 4 proactive need types, Dependency, Blockage, Risk, and Opportunity, across 10 everyday domains and 50 personal environments. Scenarios require agents to retrieve evidence from structured databases or unstructured documents before judging whether assistance is warranted and choosing its scope. On HALO-Arena, we construct HALO-8K, a dataset of 8,000 responses from 10 models across 400 scenarios, annotated by 20 experts for action level, acceptability verdicts, and human diagnosis. For rejected responses, human diagnosis covers timing, intensity, and other issues. Recorded reasoning tokens and retrieval rounds measure the effort behind these responses. The best proprietary and open-source models achieve 72.88% and 56.63% acceptability, respectively. Our central finding is that over-proactivity is the dominant observed failure mode, with Premature timing or Excessive intensity accounting for 61.20% of rejected responses. Further analyses yield complementary findings on the balance between intervention and restraint, evidence acquisition, and the calibration of timing and intensity.
Ozan Sener, Anand Avati, David Hafner, Jierui Lin, Laura Manduchi, Xiaoying Pang, Langwen Huang, Torsten Hoefler 等 10 位作者
Accurate precipitation forecasting remains elusive as physics-based models struggle to resolve mesoscale convection and machine learning often yields physically inconsistent, blurred results that consistently underestimate extreme events. We present FitzRoy, the first system that simultaneously achieves state-of-the-art predictive skill and physical realism over the continental United States. Through a reinforced diffusion framework, where a conditional discrete diffusion model is refined via reinforcement learning, FitzRoy effectively resolves the long-standing trade-off between pointwise accuracy and structural fidelity. Across lead times up to 24 hours, FitzRoy consistently matches or exceeds state-of-the-art models while preserving structural realism and capturing extreme events. Its low-latency inference further enables deployment in real-time operational settings. In case studies such as the 2025 Texas Hill Country floods and the 2026 Missouri floods, FitzRoy signals record-breaking accumulations that existing models largely missed. By bridging the gap between realism and skill, FitzRoy establishes a new frontier for high-fidelity precipitation forecasting.
Outcome-based reinforcement learning provides coarse supervision for long-horizon LLM agents: failed trajectories may contain correct steps, while successful trajectories can contain errors. Trajectory-level skills can also shift teacher preferences at unrelated steps, introducing off-target distillation signals. We introduce SCOPE, a framework for Selective Critical-Step On-Policy Self-Distillation. After a supervised warm-up, the policy analyzes its own trajectories, generates step-specific hindsight skills, and distills their guidance only at selected error steps. A bounded signed teacher-student coefficient stabilizes auxiliary updates while preserving positive reinforcement and negative correction. A success-relative pairwise length reward further encourages efficient task completion. Experiments on ALFWorld and WebShop demonstrate consistent improvements across multiple model backbones. With Qwen2.5-3B, SCOPE achieves 97.7% success on ALFWorld and 85.2% on WebShop, outperforming SEED by 5.9 and 6.3 percentage points, respectively. On unseen ALFWorld tasks, SCOPE reaches 89.6% success, improving over GRPO by 23.2 percentage points with 46.7% fewer interactions. Scaling to Qwen3-32B, SCOPE achieves 99.2% success on seen ALFWorld tasks and 94.0% on unseen tasks.
We introduce Nemotron 3.5 Super and Nemotron 3.5 Lightning, open and efficient Mixture-of-Experts Hybrid Mamba-Transformer models designed for reasoning, long-context understanding, and agentic workloads. Super is a 120B-parameter, 12B-active multimodal model, while Lightning is a smaller 30B-parameter, 3B-active text model; both designed for efficient deployment. Both models combine large-scale distillation, long-context training, and native speculative decoding, while Super further incorporates Indic continued pretraining and agentic software-engineering mid-training. We develop a comprehensive post-training pipeline spanning supervised fine-tuning and reinforcement learning (RLVR and RLHF), complemented in Super by specialized teacher training and multi- teacher on-policy distillation (MOPD). This transforms Super from a text-only base model into a multimodal agent capable of reasoning over text, images, video, and interactive tool environments. We further explore test-time scaling through parallel agent execution and self-verification. Across training and deployment, we jointly optimize capability and efficiency. Indic continued pretraining improves Indian-language benchmarks by up to 9.82 percentage points, while agentic mid-training strengthens software-engineering capabilities and robustness across agent harnesses. The post-trained Super model achieves 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra despite using approximately 4.6x fewer active parameters. Post-training MTP boosting improves speculative- decoding acceptance length by up to 21.8% under probabilistic draft sampling, while NVFP4 quantization achieves 99.62% median accuracy recovery relative to BF16. Inference optimizations enable up to 4,856 output tokens/s/GPU on GB200 while sustaining at least 50 tokens/s/user.
Shenglong Yao, Zirui Liu, Xiangchi Yuan, Wenke Lee
Recent benchmarks show that LLMs achieve strong performance in both function-level binary understanding and end-to-end binary reverse engineering tasks, but whether LLMs can reason over interprocedural semantics in compiled binaries remains unclear. Function-level benchmarks offer limited visibility into cross-function relationships, while whole-program benchmarks typically evaluate final outcomes of tasks that may be solvable from local clues alone, making it difficult to disentangle interprocedural reasoning from exploration, tool use, execution, and retrieval. To address this limitation, we introduce ICALL-BENCH, a benchmark that directly evaluates whether LLMs and agents can understand and reason over interprocedural semantics in binary programs, using indirect-call target recovery as a measurable probe. ICALL-BENCH covers 50 complete real-world C/C++ programs spanning different complexity levels and multiple optimization configurations, with an average of 74,746 lines of code and 941 functions per program. We further introduce three analysis regimes spanning model-only reasoning, static tool-augmented agents, and Internet-enabled open-world agents to examine how intrinsic reasoning, active information acquisition, and external knowledge each contribute to interprocedural binary analysis. We evaluate five recent LLMs and their corresponding agents. Our results reveal a wide separation across current models and agents, with program-level verifiable success rates ranging from 23% to 81%. They also show that greater information access does not consistently lead to better results, with particularly limited gains on the hardest cases, where deeper interprocedural reasoning remains the primary bottleneck. Uneven performance across models and frequent verifiable failures even under an optimistic evaluation indicate that robust binary interprocedural reasoning remains an open challenge.
Zheng Lian, Fan Zhang, Lan Chen, Yazhou Zhang, Rui Liu, Jinyang Wu, Haoyu Chen, Xiaobai Li 等 11 位作者
Open-Vocabulary Multimodal Emotion Recognition (OV-MER) aims to predict emotions without being constrained by predefined label spaces, thereby enabling fine-grained emotion understanding. Unlike traditional discriminative methods, OV-MER leverages generative models to capture the full spectrum of emotions and employs emotion wheels (EWs) for metric calculation. Previous approaches primarily rely on token-level loss during training. However, this objective is misaligned with the metrics used in OV-MER, and these metrics cannot be directly optimized via gradient backpropagation. To address this limitation, we turn our attention to reinforcement learning, as this strategy can optimize non-differentiable objectives. We term this framework AffectGPT-RL. Furthermore, we conduct extensive experiments to elucidate the role of reinforcement learning in this task, revealing the necessity of the reasoning process, the impact of different rewards, and the generalizability to other emotion tasks such as sentiment analysis and basic emotion recognition. Experimental results demonstrate that AffectGPT-RL yields significant performance improvements on OV-MER. Beyond this task, we also achieve remarkable performance gains on basic emotion recognition, attaining state-of-the-art results on MER-UniBench. To the best of our knowledge, this is the pioneering work exploring the role of reinforcement learning in OV-MER, providing valuable guidance for subsequent researchers. Our code is provided at: https://github.com/zeroQiaoba/AffectGPT/tree/master/AffectGPT-RL.
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.
Qingrui Li, Haowei Lou, Chengkai Huang, Quan Z. Sheng, Lina Yao
Pretrained audio encoders are increasingly used to provide content representations for music recommendation, but their downstream value can change substantially with the recommendation interface through which they are used. We study this effect across six pretrained audio encoders and three representative settings: content-based retrieval, sequential recommendation, and Semantic-ID-based generative recommendation. Using matched random-feature and ID-only controls, we distinguish absolute recommendation performance from gains over content-free inputs. Across two music datasets and a full-catalog evaluation, music-pretrained and audio-text-aligned encoders are generally stronger when their pretrained geometry is used directly or converted into Semantic IDs, while these differences become much smaller after interaction-based adaptation in SASRec. We further examine how audio representations should be discretized for generative recommendation by varying residual-quantization width, depth, and retained layers. At epoch 10, depth one performs best at every tested width, while larger codebooks generally help at depth three; the best single-layer subset also outperforms the full three-layer identifier in every evaluated comparison. For CLAP-Music, the third-layer identifier ranks higher under both original and shuffled assignments, whereas the full three-layer identifier gains more from the original assignment. Better reconstruction likewise does not ensure better ranking. Overall, our results show that the utility of pretrained audio representations is interface-dependent and that Semantic-ID quality should be assessed by downstream ranking rather than representation fidelity alone.
In communication networks, network coding allows intermediate nodes to combine messages, achieving transmission rates unattainable by routing alone in many multicast settings. In multiple unicast, where each source sends a message only to its prescribed destination, a central question is under what conditions fractional routing suffices to attain the transmission rates supported by network coding. We prove that, for any given collection of multiple-unicast demands in an undirected network, if the maximum flow between every pair of distinct destinations is at least the total demand rate, then the following are equivalent in feasibility: multiple-unicast coding, multiple-unicast fractional routing, and simultaneous fractional routing of all source rates to any fixed destination. Verifying this pairwise maximum-flow condition and, when it holds, deciding feasibility of the given multiple-unicast demands require at most as many maximum-flow computations as there are destinations. As a direct corollary, we prove the Langberg--Médard conjecture: if a directed network supports coded multicast of independent source messages at a common rate to every destination, then removing arc directions without changing capacities permits simultaneous fractional routing at that rate for all prescribed source--destination pairs.
Streaming speech-to-speech translation requires timely decisions about when partial input is sufficient for accurate translation. We present S3-S2ST, a model that learns to translate and speak from semantically aligned streaming supervision. By converting utterance-level speech pairs into incremental training examples, we teach a learned policy when to commit translations. Its adjustable threshold controls the quality--latency tradeoff on a fixed 640\,ms input grid. S3-S2ST delivers competitive English--Chinese translation on FLEURS and leads the evaluated streaming systems in translation quality on the internal 9-pair test set. On RealSI, adapting the threshold to buffered speech further improves quality over a low fixed threshold with little change in first-audio latency. We plan to release the model and code.
Wenjun Wang, Yuchen Fang, Fengrui Liu, Zibo Liang, Kai Zheng
Large Language Model (LLM) agents have shown remarkable capabilities in autonomous decision-making by generating sequential trajectories of states, actions, and observations. However, in complex, long-horizon tasks, these agents frequently suffer from compounding errors and struggle to recover from failures. Despite advances in experience memory, extracting reliable and transferable correction knowledge from historical trajectories remains challenging. Existing methods, whether based on test-time reflection or offline memory construction, largely rely on LLMs to diagnose failures and formulate corrective strategies, which can produce brittle or task-specific insights that are difficult to generalize. To address this challenge, we propose Experience Memory Graph (EMG), a framework that reformulates agent failure recovery as a graph matching problem. During offline construction on the training set, we convert both failed exploration trajectories and successful expert trajectories into directed action decision graphs. By matching these graphs, we extract common subgraphs and graph edit paths that explicitly indicate how to correct failures, and store them in a memory graph with intra-task nodes and cross-task edges. At test time, EMG retrieves relevant insights and guides the agent in a single test-time attempt without additional reflection-and-replay rounds. Experiments on ALFWorld and ScienceWorld demonstrate consistent improvements over the evaluated memory construction and competitive baselines across seen and unseen splits.
Jia-Nan Wang, Zixun Huang, Kairui Li, Zilin Wang, Lei Wu
We study when and how momentum improves large-batch training in the one-pass regime, using power-law kernel regression as a tractable setting. We first characterize stochastic stability through the \emph{critical learning rate}, defined as the largest learning rate for stable training, and obtain $η_{\mathrm{SGD}}^{\mathrm{crit}}\eqsim 1$, $η_{\mathrm{Polyak}}^{\mathrm{crit}}\eqsim \min\{1,B(1-β)\}$, and $η_{\mathrm{Nesterov}}^{\mathrm{crit}}\eqsim \min\{1,B^q(1-β)\}$, where $B$ is the batch size, $β\in[0,1)$ is the momentum factor, and \(q>1\) is the capacity exponent governing spectral decay. Within the stable region, we derive scaling laws for the full risk dynamics, capturing the progression from an early transient through power-law decay to a noise floor. We then optimize the final-step risk under a fixed data budget, obtaining a three-regime phase diagram that reveals how the benefits of momentum change with batch size. Notably, Polyak enlarges the critical batch size, the largest batch size preserving the optimal small-batch data-scaling exponent, thereby allowing fewer sequential updates without sacrificing data efficiency. In contrast, Nesterov improves the optimal data-scaling exponent in the large-batch regime through look-ahead-induced suppression of noise amplification. Experiments on PLK regression quantitatively support the theoretical scaling predictions, while neural-network experiments provide qualitative evidence for the predicted batch-size dependence and stability behavior.
DeepInfer-Mac runs the 750B-class DeepSeek-V4.1-Flash model on a MacBook Air with 24 GB of unified memory. A mixed low-bit conversion occupies 338 GiB on the internal SSD. About 7.6 GiB of weights remain resident; routed experts and Engram rows are loaded when needed. This separation makes the model executable, but it leaves each generated token waiting for a succession of storage reads. We organize expert weights so that reads fill reusable GPU buffers directly, and couple cache allocation with the lifetime of those buffers. In end-to-end runs, the system generates English, Chinese and code at 0.21-0.72 tokens per second, with a largest sampled process footprint of 16.73 GiB across the tested requests. Matched decode comparisons show a 21.0% throughput gain from next-layer prefetching over resident-expert overlap alone, at the cost of additional reads. Longer prompts remain costly. We describe the design, provide the implementation and measurement scripts, and identify the limits of this operating point: a model far larger than memory can run locally, while storage and prefill still govern the user's wait.
Large-scale linear programs with millions of variables and constraints strain time and memory budgets. First-order methods such as the primal-dual hybrid gradient scale to this regime because they avoid matrix factorizations, but they require many iterations. Learned accelerators can reduce these iterations, yet fewer iterations do not guarantee a shorter solve. Learning inference also takes time, and running the accelerator longer trades solver time for inference time. Existing unrolled accelerators handle this trade-off poorly: each layer processes every variable and constraint, the predicted iterates still need solver refinement, and every instance runs the same layers. We therefore ask: What neural accelerator design achieves end-to-end speedups on large-scale linear programs across sizes and target accuracies? We introduce AHEAD, an accelerator that predicts step sizes for first-order linear programming solvers while keeping their update rules. Its weight-shared predictor reads compact solver statistics and rescales the primal and dual step sizes, so each warm-start iteration stays inexpensive and a short warm-start phase hands the solver a warm start. A learned handoff policy sets the number of warm-start iterations for each instance and target accuracy, after which the solver finishes under its own termination criterion. On packing linear programs with up to $10^7$ variables, AHEAD achieves $6.5$--$11.6\times$ end-to-end speedups over Gurobi's first-order solver from a cold start, including on instances of unseen sizes and distributions. The gains extend to PageRank linear programs and hold across difficulty levels and target accuracies.
Under a fixed embedding-storage budget, recommenders can use a small full-precision model, a distilled student, or a higher-rank low-precision model. We asked which choice preserves rare-item recall and whether teacher fidelity reflects it in seven dataset-backbone cells: matrix factorization on Gowalla, Last.fm, Amazon Video Games, Steam, Goodbooks-10k and MovieLens-20M, and LightGCN on Amazon Video Games. Seven candidates fit in 168 bytes per user or item row: 40-dimensional Bayesian personalized ranking (BPR), two 40-dimensional unbiased knowledge distillation (UnKD) students, and factorizations of a 300-dimensional teacher's scores at ranks 40, 80, 155 and 300, stored in 32, 16, 8 and 4 bits. Each test set was evaluated once after fixing candidates and confirmatory analyses. The rank-300 4-bit factorization used 13% of teacher storage and retained 95-103% of its tail recall@20 and 93-97% of its NDCG@20 in every cell. The rank-40 factorization had the lowest tail recall in every cell; the 4-bit factorization had 1.2-12.7 times its tail recall and 0.99-1.46 times its NDCG. UnKD students retained 56-102% of teacher tail recall. In all ten prespecified comparisons across five prospective-expansion cells, the 4-bit factorization had higher tail recall, with simultaneous 95% intervals excluding zero. Teacher-hit retention, the share of the teacher's relevant tail hits also found by a model, ordered 23 of 147 candidate pairs differently from tail recall and identified the wrong tail-recall leader in 3 of 7 cells: models can find relevant tail items the teacher misses. In these cells, spending bytes on rank rather than precision preserved rare-item recall at 7.5 times the arithmetic per score. Compressed recommenders should be judged by total tail recall, not teacher fidelity alone.
Zehao Zhao, Shangjian Yin, Rui Liu, Chongyang Bai, Mark Ibrahim, Kavosh Asadi, Yuchen Lu, Hang Cui 等 9 位作者
Test-time updating lets a language model change persistent state after training---such as model weights, adapter parameters, or a memory module---before or while answering queries. Existing evaluations often compare systems using update count or frequency and performance measured immediately afterward. These summaries omit what is updated, where, when, and in what order. The same total can therefore hide an important difference: one system may answer after one update while another answers after two. We introduce the Control Map, an evaluation framework that records the committed updates preceding each scored answer and separately evaluates behavioral changes, immediate learning (acquisition), later recall (retention), preservation of unrelated behavior (locality), and performance in new settings. We prove when counts alone determine final state and show, under stated assumptions, how differences between pre-answer histories limit differences in average scores. Using ControlMap, our experiments reveal three insights. First, update scheduling matters: at a 760M-parameter checkpoint, 24 schedules with the same update count yield 19.7%--34.2% accuracy. Second, learning an edit does not guarantee that it lasts or remains contained: all six selected CounterFact LoRA settings learn the new information, while ROME retains edits at all 13 operating points but fails locality at each. Third, no editor performs best consistently across models, benchmarks, and tests. Evaluations should therefore report update histories and assess acquisition, retention, and locality separately.
Decision models enable software to select from predefined actions using natural language inputs. However, the external evidence that supplies facts for these decisions may also contain instructions that attempt to override the application's policy. We propose TrustJev, a framework combining protected policy scoring with bounded evidence corrections. A single Transformer backbone computes policy scores and a correction budget from trusted policy and candidates, then bounds evidence corrections by that budget. Paired supervision encourages stable predictions under instruction insertions that preserve task facts and changed preferences when edits change the correct answer. Authority supervision accommodates evidence when the policy alone cannot determine the decision. We evaluate a 0.6B implementation against 14 baselines on three datasets covering authentication, instruction hierarchy, and executable authorization rules. TrustJev achieves the highest mean Balance, the harmonic mean of accuracy on non-attack and attack inputs, exceeding the strongest baseline on each by 0.57-1.05 percentage points. In the authorization benchmark, it correctly answers all 32 attacks per source in a fixed pool for 95.77% of sources, compared with 80.98% for the strongest comparator. TrustJev maintains comparable inference memory and median latency, with additional overhead primarily during training.
The sampling procedures of diffusion and flow-matching models involve hundreds of denoising steps incurring substantial inference time, which has motivated a growing body of work on training schemes for few-step and one-step generation. Existing approaches primarily seek to align the generation trajectory of few-step and base models, yet their training objectives only involve discrepancies at a single point in the trajectory, encouraging only local \emph{point-level} alignment. While recent methods such as \emph{Shortcut-CSL} introduce cumulative objective that enforce alignment over the entire trajectory, this is suboptimal as they do not directly enforce alignment between the distributions induced by the two models throughout the generation process. We propose \emph{Cumulative Marginal Matching} (CMM), a principled objective for few-step generation: by viewing flow-matching as a sequence of intermediate marginal distributions connecting the prior and data distributions, CMM directly optimizes the cumulative discrepancy between the base and few-step marginal distributions over the entire sequence. Specifically, we formulate this discrepancy using the Wasserstein-2 distance. By exploiting the pushforward representation of flow-matching, we transform an intractable cumulative Wasserstein objective into a tractable surrogate that can be optimized using samples from the intermediate marginal, which are readily generated through the conditional interpolation process of flow-matching. Furthermore, we show that the existing \emph{Shortcut-CSL} objective arises as a first-order approximation of that of the CMM, establishing a theoretical connection between the two objectives. Experiments on standard benchmarks demonstrate improvement in generation quality over \emph{Shortcut-CSL} and other state-of-the-art one- and few-step baselines under identical training budgets.
Jiarui Liu, Terry Jingchen Zhang Vector Institute, Ryan Faulkner DeepMind, Xuanqiang Angelo Huang, Vilém Zouhar Department of Computer Science, ETHZ - ETH Zurich, Dominik Glandorf EPFL - EPF Lausanne, Isabel Dahlgren, Rishit Dagli University of Toronto 等 21 位作者
Expert writing feedback from experienced researchers is critical for early-career scholars to improve their manuscripts, yet high-quality feedback often remains scarce because reviewing research papers is labor-intensive. Emerging AI-powered writing assistants largely focus on grammar fixes or simulating peer review with final scores, yet they fall short of providing concrete, actionable suggestions that help students improve their papers during drafting. We present PaperMentor, a human-centered writing assistant system that delivers actionable suggestions as Overleaf-native inline comments while leaving the actual writing entirely to human authors. PaperMentor integrates an expert skill library carefully curated from established researchers’ writing advice with 12 specialized agents covering different aspects of paper writing, such as formatting compliance, phrasing accuracy, and terminology consistency. In a user study (n=14), 90.6% of the generated comments were rated actionable and 67.5% were rated valid, significantly outperforming a GPT-5.2 baseline without the skill library. We release PaperMentor as open source for public use.
尚无 AI 解读，展示原文摘要
cs.CL · cs.MA
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) · 2026 · 正式发表
Jiawei Cao, Jie Ouyang University of Science and Technology of China, Mingyue Cheng University of Science and Technology of China, Zhaomeng Zhou, Chunli Liu Hefei University of Technology, Yupeng Li University of Science and Technology of China, Zirui Liu, Shijin Wang State Key Laboratory of Cognitive Intelligence
Retrieval-Augmented Generation (RAG) is a mainstream approach to mitigating hallucinations in Large Language Models (LLMs), yet in dynamic real-world scenarios, such as weather forecasting or evolving news events, existing retrievers suffer from both temporal-semantic misalignment and outdated-document interference. To address this, we propose Relevance Recency Retrieval (Re^3), a novel framework that mitigates temporal hallucinations via two core components: a Time-Aware Dual Relevance Encoder that embeds heterogeneous temporal signals into the semantic space to ensure retrieval fidelity, and a Conflict-Aware Recency Filter that performs listwise arbitration to identify and suppress obsolete factual versions. To rigorously evaluate this setting, we introduce Re^2 Bench, a large-scale benchmark comprising over 1.3 million instances designed to assess system robustness in realistic environments where temporal constraints and conflicting factual versions coexist. Experiments on three public benchmarks and Re^2 Bench demonstrate that Re^3 consistently outperforms the strongest baselines by an average of 9.7% in generation accuracy, with gains of up to 25.2% on challenging dynamic tasks, while demonstrating robustness across diverse RAG settings.
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
Weikang Shi, Aldrich Yu Information Engineering, Chinese University of Hong Kong, Rongyao Fang The Chinese University of Hong Kong, Houxing Ren Beijing University of Aeronautics and Astronautics, Ke Wang The Chinese University of Hong Kong, Aojun Zhou, Changyao Tian, Xinyu Fu Huawei Technologies Ltd. 等 14 位作者
While Large Language Models (LLMs) have excelled in textual reasoning, they struggle with mathematical domains like geometry that intrinsically rely on visual aids. Existing approaches to Visual Chain-of-Thought (VCoT) are often limited by rigid external tools or fail to generate the high-fidelity, strategically-timed diagrams necessary for complex problem-solving. To bridge this gap, we introduce MathCanvas, a comprehensive framework designed to endow unified Large Multimodal Models (LMMs) with intrinsic VCoT capabilities for mathematics. Our approach consists of two phases. First, a Visual Manipulation stage pre-trains the model on a novel 15.2M-pair corpus, comprising 10M caption-to-diagram pairs (MathCanvas-Imagen) and 5.2M step-by-step editing trajectories (MathCanvas-Edit), to master diagram generation and editing. Second, a Strategic Visual-Aided Reasoning stage fine-tunes the model on MathCanvas-Instruct, a new 219K-example dataset of interleaved visual-textual reasoning paths, teaching it when and how to leverage visual aids. To facilitate rigorous evaluation, we introduce MathCanvas-Bench, a challenging benchmark with 3K problems that require models to produce interleaved visual-textual solutions. Our model, BAGEL-Canvas, trained under this framework, achieves an 86% relative improvement over strong LMM baselines on MathCanvas-Bench, demonstrating excellent generalization to other public math benchmarks. Our work provides a complete toolkit—framework, datasets, and benchmark—to unlock complex, human-like visual reasoning in LMMs.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Ning Li, Zheng Zhang, Zhenya Huang University of Science and Technology of China, Rui Li, Yi Zhan, Yinbo Luo, Qi Liu University of Science and Technology of China, Enhong Chen University of Science and Technology of China
The rapid advancement of large language models (LLMs) has driven the deployment of LLM-based AI tutors on online learning platforms. This widespread adoption highlights an urgent need for systematic benchmarks to evaluate their tutoring capabilities. However, existing evaluations predominantly focus on isolated, short-term interactions, overlooking the inherently long-term nature of learning. To bridge this gap, we introduce LongTutor, a benchmark for long-term personalized tutoring grounded in formative assessment theory. Built from expert-annotated real-world learning logs, LongTutor evaluates LLMs across three progressive tasks: historical evidence acquisition, knowledge state diagnosis, and adaptive teaching action. Our experiments reveal a critical capability mismatch: while LLMs excel at evidence acquisition, they struggle to effectively leverage long-term history for accurate diagnosis and adaptive teaching. To enable scalable benchmark expansion, we further propose an automated generator–verifier pipeline, paving the way toward truly long-term AI tutoring systems.
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
Large Language Model-based Multi-Agent Systems (MAS) have demonstrated remarkable capabilities in complex tasks. However, manually designing optimal communication topologies is labor-intensive, while automated expansion methods often result in bloated structures with redundant agents, leading to excessive token consumption. To address this problem, we introduce AgentSlimming, a plug-and-play compression framework for graph-structured multi-agent workflows. Motivated by the AgentPruner and AgentQuant in neural networks, AgentSlimming compresses workflows by firstly estimate the importance score of each agent with a hybrid mechanism, and then removing redundant agents or replacing them with low-cost ones, where each operation is then validated with a baseline-anchored acceptance rule to prevent performance collapse. Experiments show that AgentSlimming reduces average token cost by up to 78.9% with negligible performance degradation, and even sometimes improves accuracy, achieving a strong Pareto-optimal trade-off between cost and quality.
尚无 AI 解读，展示原文摘要
cs.CL · cs.MA
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Zheng Zhang, Qi Liu University of Science and Technology of China, Siyuan Liang Nanyang Technological University, Ning Li, Zirui Hu Nanyang Technological University, Weibo Gao University of Science and Technology of China, Rui Li, Zhenya Huang University of Science and Technology of China 等 11 位作者
Large language models (LLMs) have achieved remarkable performance across diverse tasks, largely driven by large-scale pretraining. However, this data abundance introduces test data contamination, where benchmark datasets overlap with pretraining corpora, undermining the reliability of model evaluation by confounding memorization with genuine generalization. To mitigate this issue, existing training data detectors attempt to identify clean (unseen) samples from contaminated test sets, but often suffer from residual contamination due to the black-box nature of LLMs. As a result, contaminated data may be mistakenly retained, leading to unreliable evaluation.To address this challenge, we propose FTD (FDR-controlled Training Data detection), a principled framework that detects and filters contaminated evaluation data while providing a statistical guarantee: the proportion of contaminated samples mistakenly retained as clean, the false discovery rate (FDR), is provably controlled below a user-specified threshold. FTD combines multiple complementary detectors via an adaptive weighting strategy, and we theoretically show it achieves high statistical power under valid FDR control. Extensive experiments on real-world benchmarks demonstrate that FTD significantly reduces residual contamination compared to existing methods while preserving evaluation consistency.
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
Wenrui Liao, Weihong Du, Yi Li, Hongru Liang Sichuan University, Wenqiang Lei Sichuan University
Automated schedule generation for multitask from natural language descriptions has huge potential in modern industry. While classic methods bypass language complexities by using pre-formatted matrices, and recent LLM+solver approaches introduce new fragilities by relying on solver-specific code generation. This raises critical questions: Can large language models (LLMs) solve this NL \Rightarrow Schedule task end-to-end well(RQ1)? If the answer is “no”, where do they fall short(RQ2)? And how can their capabilities be enhanced (RQ3)? To answer these questions, we introduce NL \Rightarrow Schedule, the first benchmark for this task, equipped with a dataset of 240 description-schedule pairs constructed from real-world materials and a rigorous evaluation suite. Our evaluation of nine state-of-the-art LLMs reveals the limitations of different LLMs in procedure grounding and the strengths of advanced LLMs in global planning via local analysis. To address these shortcomings, we propose Mans, a novel multi-agent framework. Extensive experiments show that Mans achieves more robust performance comparable to six state-of-the-art LLM+solver methods. We hope NL \Rightarrow Schedule and Mans will serve as a solid foundation for automatic scheduling.
尚无 AI 解读，展示原文摘要
cs.CL · cs.MA
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Weidong Tang, Jierui Li Department of Computer Science and Technology, Tsinghua University and Xidian University, Yueling Hou Xi'an University of Electronic Science and Technology, Zihan Mei, Can Zhang, Xinyan Wan, Zhiyuan Liang University of Michigan - Ann Arbor, Pengfei Zhou 等 10 位作者
True general intelligence requires not only a model of the physical world but also a social world model: the capacity to infer how individual mental states interact and crystallize into group-level outcomes. Despite notable progress in individual-level Theory of Mind (ToM) reasoning, existing multimodal large language models systematically fail at this: collective behavior emerges non-linearly from social tensions, conformity dynamics, and structural constraints, and cannot be recovered by summing individual intentions. We present GroupToM-Bench, the first multimodal benchmark for group-level ToM, built around a causal chain spanning micro-level BDI states (belief, desire, intention), meso-level group tension and structural constraints, and macro-level outcome prediction and mechanistic attribution. To probe this full arc, we develop a seven-level cognitive audit framework. Experiments reveal that frontier models perform significantly below human levels, exposing fundamental blind spots in modeling social structures and nonlinear collective behavior.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表