Kun Gu, De Cheng, Zhipeng Xu, Lingfeng He, Di Xu, Nannan Wang
Continual learning (CL) aims to adapt models to new tasks while preserving previously learned knowledge. Recently, Low-Rank Adaptation (LoRA), a representative Parameter-Efficient Fine-Tuning (PEFT) method, has gained increasing attention in CL due to its efficiency and scalability. Several LoRA-based merging methods maintain a single inference model by scaling the newly learned LoRA update with one coefficient and merging it into the previous weights after each task. However, they usually control the whole update with a shared scalar factor, although different update elements may affect previous and current tasks in different ways. To address this issue, we propose Merging Strength Modulation via Joint Loss Estimation (M2LE). It consists of Hessian Information-guided Adaptive Merging (HIM) and Task Simulation Perturbation for Merging (TSP). Instead of searching for a global merge coefficient, HIM directly solves for the target merged update under a joint objective over previous and current tasks. Solving this objective gives an analytic solution that naturally takes the form of element-wise modulation over the current task LoRA update. To improve the stability of the learned update under subsequent merging, we introduce TSP which simulates different tasks interference. Experiments on multiple benchmarks show that M2LE consistently outperforms existing PEFT-based continual learning methods. Code is available at https://github.com/bbdeeer/M2LE_NeurIPS2026.
The kernel of the Lean proof assistant checks that a proof establishes its statement. Whether the statement says what was intended is left to human reviewers. This is the old problem of validating a specification, and it grows more pressing as language models write formal statements in bulk and agents edit formal developments continuously. We present ProofPatch, a tool that helps reviewers judge Lean statements and keeps their judgments valid while the development changes. A reviewer can ask the kernel targeted questions about a statement, such as whether a hypothesis is needed or whether a counterexample exists. Each decision is stored with fingerprints of the statement as Lean elaborated it and of the definitions it depends on, and any declaration can be pinned so that a patch changing it is blocked, whether a person or an agent built on a large language model (LLM) wrote the patch. Working on elaborated terms rather than source text is the central design choice: notation and formatting changes leave a decision intact, and a changed definition reaches every theorem that uses it. On 229 pairs of consecutive revisions of Mathlib, Lean's mathematics library, ProofPatch sends 118 of 12,898 synthetic decisions back for review. Every text-based alternative misses at least 38 of these or sends back at least 13 times as many. On 1,527 labeled statements, kernel evidence flags unfaithful ones with precision 0.88 to 0.96, and in 344 runs of an LLM agent no change to a pinned statement got through.
Determining the neutral molecular formula of a precursor ion from its tandem mass (MS/MS) spectrum is the first step in identifying an unknown small molecule. Existing methods trade recall, whether the correct formula is among the candidates, against the cost of examining each candidate. Methods that examine each candidate in depth must first shorten the list, by limiting element counts or truncating it, and that is where the correct formula is lost. End-to-end regressors are fast but guarantee neither validity nor recall. We present FACET, which separates deciding which formulas to consider from deciding which one to choose. A feasibility automaton builds a formula one element count at a time and admits exactly the formulas that fit the precursor mass and satisfy the ring-and-double-bond and valence constraints. Run as an exhaustive search, it enumerates the complete feasible set, so every candidate is valid and recall is fixed by the chemical constraints rather than by a trained model. A small neural scoring head then ranks every feasible formula from one pass of a pretrained spectrum encoder, so no candidate is discarded. On the NPLIB1 benchmark FACET reaches top-1 accuracies of $0.859$ with the adduct known and $0.729$ with it unknown, compared with $0.766$ and $0.660$ for the strongest established method, and its recall is the highest. On the external CASMI 2022 set its recall is $1.000$, and with the adduct unknown it is the most accurate of the methods compared.
External skills provide language-model agents with reusable procedural knowledge, yet how models use such knowledge after it is supplied remains poorly understood. We identify Procedural Exposure Bias (PEB), the tendency of language models to increasingly favor an action as its associated procedural pattern is repeatedly exposed in a skill, even when the interaction state and action relevance remain unchanged. We find that PEB strongly depends on how procedural knowledge is represented: presenting the same content visually substantially attenuates exposure-induced shifts in action preference relative to textual presentation. Further analysis shows that this attenuation primarily arises from visual encoding rather than structured layout. Although repeated procedural signals remain recoverable from Visual representations, they exert a weaker influence on downstream action prediction, distinguishing signal accessibility from its use as action evidence. This effect also extends to executable action selection, where increasing exposure more strongly favors the corresponding action under Text than Visual. Layer-wise analyses show that exposure-dependent action evidence becomes more pronounced toward the model output, with stronger evidence under Text, while activation interventions reveal architecture-specific roles of internal components in transmitting or counteracting the preference shift. Complete-agent evaluations across ALFWorld, ScienceWorld, and WebShop further show that Visual skills achieve higher macro-average task success than Text skills for all four evaluated models. These findings establish skill representation as an important factor governing how procedural knowledge is translated into agent decisions. All code is available at https://github.com/NEUIR/Vis-Skill.
Reza Yousefi Maragheh, Jayesh Uddhav Kudase, Aysenur Inan, Ramin Giahi, Kai Zhao, Parth Parikh, Jianpeng Xu, Jason Cho 等 10 位作者
Evaluating natural language explanations in recommender systems is essential for fostering user trust, transparency, and engagement. However, existing evaluation approaches like human evaluations, while accurate, are resource-intensive and impractical at the scale required by modern recommendation platforms. Also, automated methods using single-agent LLMs suffer from prompt sensitivity and inconsistent outputs. To address these challenges, we propose MAREval, a structured multi-agent framework for evaluating recommendation explanations using large language models. MAREval orchestrates (i) a planner agent that uses a novel Chain of Debate (CoD) prompting strategy to coordinate agent roles and enforce logically consistent evaluation plans; (ii) a moderator agent that regulates discussions by mitigating prompt drift; and (iii) an arbitrator agent that aggregates outputs from multiple evaluation rounds and (iv) a Monte Carlo sampling method, improving robustness and alignment with human judgment. We conduct comprehensive evaluations on both public (TopicalChat) and proprietary recommendation datasets, demonstrating that MAREval outperforms state-of-the-art baselines. Comprehensive experiments on a public benchmark and a proprietary e-commerce dataset show that MAREval improves alignment with human judgments over strong singleand multi-agent baselines. Stability analyses indicate substantially lower variability across repeated trials. In a large human-annotation gate, MAREval meets production quality thresholds where prior evaluators fall short, and online A/B testing demonstrates statistically significant improvements in engagement and revenue metrics. These results establish MAREval as a scalable and reliable solution for human-aligned evaluation of recommendation explanations in real-world systems
Sequential recommendation models typically predict a user's next action from recent session interactions, which limits personalization when short-term signals are weak or ambiguous. Rather than assuming that incorporating longer-term history and community signals is universally beneficial, we study when it actually helps. We propose BehaveRec, a three-signal framework combining a causal session transformer, a bidirectional User History Encoder over a 90-day window, and a GraphSAGE module over a user-user graph built from behavioral embedding similarity rather than item co-occurrence; a learned gate blends the three signals per user. On an industrial dataset of 137M sequences from 12.6M users, BehaveRec improves R@5 by +3.2% over SASRec, with a larger-capacity SASRec baseline (more parameters than BehaveRec) still trailing by +2.5%, ruling out capacity as the explanation. More importantly, a user-level stratification by history length reveals a clear dose-response relationship: BehaveRec's advantage is negligible below roughly 10 interactions and grows monotonically with history richness, reaching +0.017 R@5 for users with 101+ events. This pattern is corroborated across one industrial and three public datasets spanning history-rich to history-sparse regimes, where session-only models remain preferable. These headline results, and the resulting deployment threshold, are established for category-level intent prediction; a preliminary, single-run check on the full item-level vocabulary yields a directionally mixed result and is scoped as an open question rather than a claim. We use these results to derive a practical deployment rule -- roughly 20 interactions per user -- for when long-term history and behavioral-graph signals are worth their added complexity. Despite this added complexity, BehaveRec remains lightweight: it uses 5.67M parameters and trains on a single GPU in under two days.
Guoyi Zhang, Yanjin Du, Guangsheng Xu, Zhengyao Zhao, Siyang Chen, Jingjing Gao, Yuezheng Zhou, Yiyang Wu 等 19 位作者
Small uncrewed underwater vehicles are proliferating and threaten maritime security, yet evade sonar and magnetic sensors built for larger submarines. Infrared thermal wakes, generated by hydrodynamic displacement that modulates the cold-skin layer, offer a non-acoustic, non-magnetic alternative even without cooling-water discharge. Existing studies, however, rely on idealized simulations that fail to capture real-world backgrounds, offering limited guidance for target-background discrimination. We establish AVP-IR-Wake, the first large-scale, real-world, physically diverse benchmark of 161,485 frames collected over three years. We show that thermal wakes are extended structures of discrete vortices, violating the compactness assumption of conventional targets. We reveal local indistinguishability: individual wake vortices are statistically identical to wind-induced background vortices in temporal, frequency, and sparse domains, sharing low-frequency information and long-range correlations. This challenges the assumptions of compact, transient, high-frequency targets in conventional detection. Yet we establish global morphological separability: although locally inseparable, the wake as a whole is organized by vehicle trajectory, whereas background clutter is organized by wind-driven thermal fronts. This explains why representative detectors fail and shows that reliable detection must shift from local feature extraction to global trajectory-constrained physical modeling. Because the cold-skin layer varies globally and seasonally, forecasting wake observability will require integrating physical oceanography and atmospheric science, linking underwater surveillance to the changing ocean surface.
Recent advances in vision-language-action (VLA) and world-action models (WAMs) have significantly improved robotic manipulation, yet deploying robots in the open world demands a higher level of autonomy beyond executing predefined tasks. Robot agents offer a promising paradigm for this transition, integrating task discovery, contextual reasoning, physical execution, and adaptive recovery into a closed interaction loop. However, existing benchmarks largely evaluate isolated capabilities, failing to capture the complexity of open-world robot autonomy. We introduce RobotAgentBench (RAB), a comprehensive benchmark for evaluating robot agents in realistic, open-world household environments. Through an agent-assisted, human-in-the-loop real-to-sim pipeline, RAB constructs 50 diverse household layouts with interactive objects and physically grounded tasks, encompassing 90 atomic tasks, 18 short-horizon tasks, 8 long-horizon tasks, and 4 open-ended scenarios. RAB provides a multi-level evaluation of robot agent capabilities: (1) task-level assessments systematically examine five complementary dimensions: task discovery, context reasoning, manipulation, failure recovery, and robustness; (2) scenario-level evaluation assesses how agents integrate these capabilities to autonomously fulfill high-level user needs. Furthermore, RAB enables systematic comparisons of standalone VLA/WAM models and diverse robot agent architectures under a unified planning-execution-verification framework. Extensive experiments reveal substantial gaps between manipulation and autonomy, demonstrate the advantages of agentic architectures and recursive self-improvement, and highlight the remaining challenges toward open-world robot autonomy.
Untargeted tandem mass spectrometry (MS/MS) detects thousands of small molecules per biological sample, yet most go unidentified because they are absent from spectral libraries. These uncharacterized metabolites and natural products are precisely the compounds that matter for drug discovery, biomarker research, and exposomics. Computational de novo structure elucidation could close this gap, but almost all state-of-the-art methods assume the ground-truth molecular formula is known, an oracle that does not exist for genuinely novel compounds, whose formula must itself be predicted with substantial error. We present MARLIN, a de novo method that elucidates structures directly from a spectrum with no molecular formula at any stage. A self-supervised encoder predicts a molecular fingerprint from the raw peaks, and a block-diffusion language model generates candidate structures conditioned only on the fingerprint and the instrument-measured precursor mass. A mass-shell constraint, whose heavy-atom prune provably never excludes the true molecule, steers generation toward the measured mass without fixing the atom inventory, and candidates are accepted by exact parts-per-million mass agreement. A symmetric noise objective absorbs encoder error, and a candidate-diversity mechanism keeps the candidates from collapsing to a single structure. On the NPLIB1 benchmark, MARLIN reaches 16.9% Top-1 exact-match accuracy, higher than the formula-given baselines, and returns the closest candidate structures by structural distance and fingerprint similarity. It recovers the correct molecular formula as a byproduct about as often as a dedicated predictor without ever using one. MARLIN brings de novo structure elucidation to the discovery regime where the molecular formula is unavailable.
Yuewen Liu, Peng Xu, Muxi Diao, Anyi Zhang Beijing University of Posts and Telecommunications, Yang Li, Yutong Zhang
Large Language Models (LLMs) are inherently constrained by their fixed-length context windows, which limits LLMs’ ability to retain and utilize information across long-term interactions. To address this limitation, recent work has proposed external memory modules for LLMs. Using memory modules typically involves two stages: evidence retrieval and memory utilization. While prior work focuses on the architecture of memory modules and the retrieval stage, the equally critical memory utilization stage remains underexplored. Building on this, we propose MemCoRL, a two-stage alternating co-optimization reinforcement learning method. Stage 1 optimizes evidence retrieval using citation feedback and semantic accuracy from utilization as rewards. Stage 2 optimizes utilization with rewards combining semantic similarity and lexical overlap. Iterative co-optimization establishes a positive feedback loop: better retrieval improves memory utilization, which in turn refines retrieval rewards. Experimental results show our approach outperforms the leading baselines on both lexical overlap and semantic similarity metrics, confirming the co-optimization in memory retrieval and memory utilization.
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
Zheng Luo, T Pranav Kutralingam, Ogochukwu N. Okoani, Wanpeng Xu Arizona State University, Hua Wei Arizona State University, Xiyang Hu Arizona State University
Large Language Models (LLMs) are increasingly deployed as agents that invoke external tools through structured function calls. While recent work reports strong tool-calling performance under standard English-centric evaluations, the robustness of tool calling under multilingual user interactions remains underexplored. In this work, we introduce MLCL, a diagnostic benchmark, and conduct a systematic evaluation of multilingual tool calling across Chinese, Hindi, and the low-resource language Igbo. Through fine-grained error analysis, we show that many failures occur despite correct intent understanding and tool selection. We identify parameter value language mismatch as a dominant failure mode, where models generate semantically appropriate parameter values in the user’s language, violating language-invariant execution conventions. We further evaluate several inference-time system strategies and find that while these strategies substantially reduce language-induced execution errors, none of them can fully recover English-level 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
Yifan Ji, Zhipeng Xu, Zhenghao Liu Northeastern University, Zulong Chen, Qian Zhang Alibaba Group, Zhibo Yang, Junyang Lin Alibaba Group, Yu Gu 等 10 位作者
Key Information Extraction (KIE) from real-world documents remains challenging due to substantial variations in layout structures, visual quality, and task-specific information requirements. Recent Large Multimodal Models (LMMs) have shown promising potential for performing end-to-end KIE directly from document images. To enable a comprehensive and systematic evaluation across realistic and diverse application scenarios, we introduce UNIKIE-BENCH, a unified benchmark designed to rigorously evaluate the KIE capabilities of LMMs. UNIKIE-BENCH consists of two complementary tracks: a constrained-category KIE track with scenario-predefined schemas that reflect practical application needs, and an open-category KIE track that extracts any key information that is explicitly present in the document. Experiments on 15 state-of-the-art LMMs reveal substantial performance degradation under diverse schema definitions, long-tail key fields, and complex layouts, along with pronounced performance disparities across different document types and scenarios. These findings underscore persistent challenges in grounding accuracy and layout-aware reasoning for LMM-based KIE. All codes and datasets are available at https://github.com/NEUIR/UNIKIE-BENCH.
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