Effective multi-domain adaptation of large language models requires prioritizing relevant domain knowledge while minimizing interference from irrelevant domains. Existing approaches commonly use Low-Rank Adaptation (LoRA) to train a separate adapter for each domain and combine these adapters within one model. However, they either weight adapters equally regardless of the input domain or dynamically weight adapters while still activating all of them including irrelevant ones. Additionally, while efficient, their low-rank structure limits performance. In contrast, upcycling is a model expansion method that adds experts and a router to a dense Transformer layer offering larger capacity than LoRA approaches. In this paper, we propose Domain-Specialized Upcycling (DS-Upcycling) for multi-domain adaptation. We introduce a new domain-aware supervision that allows the router to effectively learn to distinguish domains by selecting the appropriate experts. During fine-tuning on each new domain, we update only the corresponding expert and router while freezing all other parameters. Experiments show that our method achieves the highest average performance across the adapted domains compared to baselines while retaining general performance. With just 256 examples per domain for router supervision, DS-Upcycling achieves 90-98% routing accuracy after adaptation compared with 43-70% for routers trained without our approach.
Alex Brandes, Haig Conti Georges Sajelian, Manthan Patel, Dominik Hollidt, Chenhao Li, Matthias Heyrman, Oliver Hausdoerfer, Manuel Kaufmann 等 14 位作者
Deploying humanoid robots in unstructured terrain remains an open problem. While classic reinforcement learning struggles with the sheer complexity of real-world interactions, more promising methods leveraging human priors remain limited to models lacking contextual awareness. The restricted motion synthesis is a direct consequence of existing dataset pipelines failing to capture human-scene sequences in challenging environments. To bridge this gap between humanoid learning and scene reconstruction, we introduce the Egocentric Human-Terrain Reconstruction (EgoHTR) dataset. We develop and open-source a reconstruction pipeline capturing 55 scene-aligned 4D human motion sequences in diverse, complex environments using a multi-sensor setup of egocentric wearables and a portable 3D scanner. The resulting dataset comprises over 150k frames, which we evaluate against motion-capture ground truth, demonstrating state-of-the-art accuracy and establishing a rigorous benchmark for human motion analysis and synthesis. Further, we leverage this data to train perceptive locomotion policies, demonstrating hardware deployment on a Unitree G1 for reconstructed reference motions. Our pipeline enables community-driven dataset extensions and factors the problem to help researchers build foundational, context-aware robots that reliably traverse uneven terrain.
Adam Deryło, Alexey Gavryushin, Daniel Cremers, Marc Pollefeys, Xi Wang
Egocentric video has emerged as one of the most scalable sources of data for learning generalizable manipulation priors. Yet, despite its growing use in pretraining robotic policies, it has been underutilized as a validation tool for physical intelligence, a key component of which is understanding where to make contact with an object to achieve an intended interaction. We present EgoContact, a large-scale egocentric dataset for evaluating the generalizability and precision of learned manipulation priors through future contact-point prediction. EgoContact contains 22.5K manually verified interaction events, each with a pre-contact observation, semantic intent, and pixel-precise contact locations. We find that current methods often struggle to predict precise interaction contacts and sometimes fail to localize the target object in diverse, in-the-wild scenes, even when allowed multiple predictions. Furthermore, we demonstrate the usefulness of EgoContact's annotated interactions for robotic manipulation: simply using EgoContact as the memory bank of an existing affordance-transfer policy improves grasp success from 14.2% to 31.2% in simulation and from 16% to 58% on a real robot.
Giuliano Martinelli, Abelardo Carlos Martinez Lorenzo, Satvik Garg, Linxi Wang, Riccardo Orlando, Piriyakorn Piriyatamwong, Samuel Paul Fraiberger
AI systems that ground their answers from retrieved documents usually need only a few relevant passages, but applications such as systematic reviews and legal discovery need nearly all of them, because their conclusions rest on the full body of evidence. These applications have traditionally searched with curated taxonomies, which can only find what their vocabulary describes. We compare taxonomy-based retrieval with an alternative neural retrieval based on sentence embeddings and followed by language model screening in ImpactAI, an evidence synthesis application built on 7,354 randomized trials. Measured against the studies included in 29 systematic reviews, the share of studies found rises from 45% to 85%, and the screening model agrees with human annotators as well as annotators agree with each other. This gain comes at a cost in speed: at the median, retrieving and screening evidence for a query takes 77 seconds, against 6 seconds for taxonomy retrieval. The results suggest finding evidence with embeddings, organizing it with taxonomy labels, and setting how many candidates to screen according to the time available for each query.
Diffusion-based denoising commonly uses Gaussian perturbations, whereas sensor noise follows diverse signal-dependent statistical laws. In this paper, we propose $p$-Bridge, a unified measurement-indexed diffusion framework for Tweedie noise. The noise exponent $p$ describes the variance--intensity relation, while accumulated measurement $ν$ controls the noise level. Together, they define a noise plane that includes Gaussian, Poisson, and Gamma models. The stochastic bridge connects measurement distributions through independent increments, giving denoising a physical interpretation as progressive measurement accumulation. Both the starting and stopping indices can be specified, allowing restoration from an observed measurement level to a prescribed higher level. Measurement superposition yields a shared reverse-mean update, paired with a deviance-based objective for learning the required conditional mean. A latent compound Poisson--Gamma construction enables sampling at intermediate exponents. A single exponent-conditioned network adapts to different Tweedie noise models without separate training for each exponent. Experiments show that exponent conditioning improves cross-exponent robustness over fixed-exponent models. Using estimated exponents, the shared network achieves robustness comparable to an oracle-selected set of five specialist models. Experiments on synthetic images, FMD fluorescence microscopy, and multi-sensor SAR images further demonstrate its effectiveness across different signal-dependent noise regimes.
Zhenzhen Ren, Xiaoqing Zhang, GuoBiao Li, Bowen Fang, Yanxi Wang, Xinpeng Zhang, Ke Han
Trajectory-based confidence estimation has shown promising performance for large language models (LLMs) by leveraging rich signals from their reasoning processes. However, we find that such confidence estimators can degrade substantially after agentic reinforcement learning (Agentic RL) as it induces substantial shifts in the reasoning trajectories. More importantly, despite these shifts, we uncover a reliability structure that is largely preserved across Base and RL models: $\textbf{more}$ $\textbf{stable}$ $\textbf{trajectories}$ $\textbf{remain}$ $\textbf{more}$ $\textbf{likely}$ $\textbf{to}$ $\textbf{produce}$ $\textbf{correct}$ $\textbf{answers}$. This preserved relationship provides a transferable, label-free reliability signal for post-RL confidence adaptation. Building on this finding, we propose $\textbf{StaR-TTA}$, a stability-guided framework for label-free test-time confidence adaptation. StaR-TTA jointly learns from correctness supervision and stability-derived reliability supervision on labeled Base trajectories, establishing the connection between trajectory stability and answer reliability. At test time, it reuses stability-induced reliability rankings from unlabeled RL trajectories to adapt confidence scores to the shifted trajectory distribution, without requiring target correctness labels. Across four QA benchmarks, StaR-TTA achieves an average AUROC of \(0.721\), with consistent improvements across unseen datasets, model families and scales, and different Agentic RL algorithms.
Shuwei Wang, Qiang Guan, Brian Qiu, Haoxiang Huang, Haoxi Wang, Chenyu Li, Bo Xu, Sen Cui
Exploration in physical control environments requires agents to identify actions whose consequences reveal previously unexplored aspects of the underlying dynamics. However, many intrinsic motivation methods rely on immediate prediction errors, state novelty, or local physical changes, which may overlook informative outcomes that emerge across extended temporal horizons. We investigate whether learning structured representations of reachable futures can provide a more effective basis for exploration. To this end, we introduce Hamiltonian Temporal Contrastive Curiosity (HTC$^2$), a framework that combines multi-horizon temporal representation learning with Hamiltonian phase-space structure. A horizon-conditioned contrastive critic learns temporal relationships from future-state pairs sampled within the same episode. The distance between positive-pair representations is then used as an intrinsic reward, encouraging the agent to explore futures that remain insufficiently organized in the learned representation. To incorporate information about physical dynamics, we additionally modulate this reward using bounded multi-step Hamiltonian variation, rather than directly rewarding changes in the Hamiltonian itself. We integrate HTC$^2$ with both Soft Actor-Critic (SAC) and Proximal Policy Optimization (PPO), and evaluate it across six DeepMind Control Suite tasks. Comparisons with ICM, RIDE, E3B, Plan2Explore, and a single-step Hamiltonian curiosity baseline are complemented by state-coverage evaluations, component ablations, and phase-space diagnostics to characterize the contributions of temporal learning and physical structure to exploratory behavior.
Plan or Retrieve First? A Scale Study of Tool Discovery Architectures
Mingfei Zhang, Amit Surana, Ziauddin Majid, Chenxi Wang, Michael Xu
AI 速览：在ToolBench上系统比较Retrieve-then-Plan与Plan-then-Retrieve两种工具发现范式，发现二者优劣随查询复杂度呈现交叉。
Tool-augmented LLM agents discover relevant APIs from large registries before planning execution, yet finite context windows prohibit injecting all tool definitions into a single prompt. The dominant approach, which we term Retrieve-then-Plan (RtP), embeds the user query and retrieves similar tools via approximate nearest-neighbor search. We study an alternative Plan-then-Retrieve (PtR) paradigm in which the LLM first generates Abstract Tool Stubs (hypothesized function signatures describing needed capabilities) which then serve as retrieval queries against the registry. Evaluating on ToolBench (46,984 unique API endpoints, 50 categories) across single-tool (G1), intra-category multi-tool (G2), and cross-category multi-tool (G3) query subsets, we find a clear complexity-dependent crossover: RtP outperforms PtR on single-tool queries while PtR outperforms RtP on multi-tool queries. This crossover holds at every retrieval budget (k=1 to 20), is robust across registry sizes from 1k to 47k APIs, and is consistent across model families (Gemini 2.5 Flash and gpt-oss-120b). We characterize a stub-hallucination failure mode (the LLM invents plausible but non-existent API names) that acts as a double-edged sword: on simple queries PtR's stubs abstract away proper-noun signals that RtP matches directly, yet on complex queries the same mechanism decomposes multi-faceted intent into functional sub-needs. We further analyze token overhead and show that PtR's extra discovery call roughly doubles inference cost, which pays off on multi-tool queries but hurts on single-tool ones, motivating complexity-adaptive routing as a practical direction, one that can be driven by a zero-cost signal already present in the pipeline (RtP's top-1 retrieval score).
Tianjian Li, Jingyu Zhang, William Jurayj, Xi Wang, Chuanyang Jin, Mehrdad Farajtabar, Eric Nalisnick, Daniel Khashabi
AI 速览：提出SelfCompact框架，让语言模型智能体在推理时自主决定何时及如何压缩上下文，以降低token成本并提升性能。
Long agent traces composed of chains of thought and tool calls accumulate stale content that anchor subsequent generations, and eventually outgrow the context window. Existing scaffolds mitigate it with fixed-interval compaction triggered at a token threshold. Such triggers pay no heed to trajectory structure, risking discard of partial results mid-derivation or mid-search. We propose SelfCompact, a scaffold that allows the model itself to decide when and how to compact. Specifically, it pairs two inference-time elements: (i) a compaction tool the model invokes to summarize the accumulated context, and (ii) a lightweight rubric specifying when to fire (a sub-task has resolved, or the trajectory is converging) and when to suppress (mid-derivation, or when stuck). Both are needed. The tool alone is unevenly used across open-weight models, often invoked at unhelpful moments or not at all; the rubric alone cannot act. Together, they elicit effective adaptive compaction without any fine-tuning or external supervision. We present empirical results on six benchmarks (competitive math and agentic search) and seven models. Our results show that SelfCompact matches or exceeds fixed-interval summarization at a fraction of the token cost, improving over a no-summarization baseline by up to 18.1 points on math and 5-9 points on agentic search at 30-70% lower per-question cost. Our results expose a meta-cognitive gap: although unprompted models cannot reliably tell when their own context is rotting, a lightweight rubric closes this gap, reframing when to compact as a capability that scaffolds can supply without training.
Xiaocheng Zhang Meituan, Xi Wang, Yifei Lu Qiantang Credit, Jianing Wang, Zhuangzhuang Ye, Mengjiao Bao Meituan, Peng Yan Meituan, Xiaohong Su
With the surge of online misinformation, Large Language Models (LLMs) and Reasoning Large Language Models (RLMs) serving as Automatic Fact-Checking (AFC) systems have emerged as a prominent paradigm for reliable, explainable verification. However, our empirical study reveals that this paradigm faces a critical risk asymmetry challenge when deployed in real-world under resource-constrained environments. While Hotspot Perception Ability (HPA), the capacity to dynamically allocate reasoning resources based on social impact, is essential to mitigate this risk, existing benchmarks lack the social metadata and evaluation framework to meet this urgent evaluation needs, thereby hindering the advancement of these AFC systems. To bridge this gap, we introduce TrendFact, the first benchmark capable of evaluating HPA and three fact-checking tasks. It consists of 7,643 curated samples sourced from trending platforms and professional datasets, with an evidence library containing 366,634 entries. To enable HPA assessment, we propose two novel metrics: the Explanation Consistency Score (ECS) to evaluate the reliability of verification reasoning, and the Hotspot Claim Perception Index (HCPI) to quantify the overall HPA of AFC systems. Extensive experiments demonstrate that existing AFC systems exhibit limited performance on TrendFact. Furthermore, our proposed FactISR framework effectively enhances HPA and computational efficiency for RLM-driven 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
Xinyue Fang, Zhiliang Tian National University of Defense Technology, Zhen Huang National University of Defense Technology, Ziyi Pan, Zhihua Wen, Xi Wang, Quntian Fang, Dongsheng Li National University of Defense Technology
Existing LLM hallucination mitigation methods, including prompt engineering and model optimization, either hardly alter models’ internal knowledge or have poor cross-domain generalization. Contrastive decoding mitigates hallucinations by using layer-wise differences in LLMs. However, prior studies only explore transformer-based models (e.g., GPT), ignoring other effective frameworks like mixture-of-experts (MoE) models. Since MoE alters the traditional transformer architecture, we conduct empirical studies to investigate whether similar layer-wise differences exist in MoEs. Our results show that they do not exist in MoE with shared experts; nevertheless, across different MoEs, higher layers exhibit distinct expert activation patterns between factual and non-factual outputs. Building on these, we propose EAACD, an expert-aware adaptive contrast decoding that uses expert differences in MoE’s higher layers to mitigate hallucinations on QA tasks. EAACD splits high-layer experts into a higher-reliability group and several lower-reliability groups based on their confidence and consistency. It contrasts the higher-reliability group’s prediction with each lower-reliability group’s prediction to calibrate the model’s original predictions. To strengthen this contrast, EAACD amplifies hallucinations from lower-reliability experts via attention and masking to provide stronger negative references. EAACD outperforms all baselines on four datasets
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
Lang Gao Mohamed bin Zayed University of Artificial Intelligence, Xuhui Li, Chenxi Wang, Mingzhe Li, Wei Liu, Zirui Song, Jinghui Zhang Mohamed bin Zayed University of Artificial Intelligence, Rui Yan Wuhan University 等 10 位作者
As large language models (LLMs) increasingly imitate personal writing styles, personalization has become a key challenge for machine-generated text (MGT) detection. Yet personalized MGT detection remains largely underexplored. In this work, we introduce StyloBench, the first benchmark for evaluating detector robustness under personalization, built from literary and blog texts paired with their LLM-generated imitations. Experiments across diverse detectors show pronounced performance instability under personalization, with frequent inversions relative to general-domain behavior. To better understand this limitation, we conduct an in-depth analysis and attribute it to a feature-inversion trap, i.e., features that are effective for separating human-written text (HWT) from MGT in general flip their effect in personalized contexts, ultimately misleading detectors. Motivated by this, we propose StyloCheck, a diagnostic framework for predicting detector robustness under personalization. StyloCheck identifies the inverted features and quantifies detector dependence using perturbed texts pronounced in the features. In our experiments, StyloCheck predicts both the direction and magnitude of cross-domain performance shifts with an 85% correlation to actual outcomes. We hope this work will raise awareness of the structural risks introduced by personalization and motivate more robust approaches to personalized MGT detection. The code is available at: https://github.com/mbzuai-nlp/Personalized_MGT_Detect
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
Xi Wang, Songlei Jian National University of Defense Technology, Shasha Li National University of Defense Technology, Xiaopeng Li National University of Defense Technology, Zhaoye Li, Bin Ji National University of Defense Technology, Baosheng Wang, Jie Yu
Despite extensive safety alignment, Large Language Models (LLMs) often fail against jailbreak attacks. While machine unlearning has emerged as a promising defense by erasing specific harmful parameters, current methods remain vulnerable to diverse jailbreaks. We first conduct an empirical study and discover that this failure mechanism is caused by jailbreaks primarily activating non-erased parameters in the intermediate layers. Further, by probing the underlying mechanism through which these circumvented parameters reassemble into the prohibited output, we verify the persistent existence of dynamic jailbreak paths and show that the inability to rectify them constitutes the fundamental gap in existing unlearning defenses. To bridge this gap, we propose Jailbreak Path Unlearning (JPU), which is the first to rectify dynamic jailbreak paths towards safety anchors by dynamically mining on-policy adversarial samples to expose vulnerabilities and identify jailbreak paths. Extensive experiments demonstrate that JPU significantly enhances jailbreak resistance against dynamic attacks while preserving the model’s utility.
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
Peiyang Liu, Zhirui Chen, Xi Wang Px Securities, Di Liang, Youru Li Beijing University of Technology, Zhi Cai Beijing University of Technology, Wei Ye Peking University
Monte Carlo Tree Search (MCTS) has been widely used for automated reasoning data exploration, but current supervision extraction methods remain inefficient. Standard approaches retain only the single highest-reward trajectory, discarding the comparative signals present in the many explored paths. Here we introduce Contrastive Reasoning Path Synthesis (CRPS), a framework that transforms supervision extraction from a filtering process into a synthesis procedure. CRPS uses a structured reflective process to analyze the differences between high- and low-quality search trajectories, extracting explicit information about strategic pivots and local failure modes. These insights guide the synthesis of reasoning chains that incorporate success patterns while avoiding identified pitfalls. We show empirically that models fine-tuned on just 60K CRPS-synthesized examples match or exceed the performance of baselines trained on 590K examples derived from standard rejection sampling, a 20\times reduction in dataset size. Furthermore, CRPS improves generalization on out-of-domain benchmarks, demonstrating that learning from the contrast between success and failure produces more transferable reasoning capabilities than learning from success alone.
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
Daojun Chen, Xi Wang University of Sheffield, Shenyuan Ren Beijing Jiaotong University, Qingzhi Ma Huatai Securities, Pengpeng Zhao Soochow University, An Liu Soochow University
While Large Language Models have significantly advanced Text2SQL generation, a critical semantic gap persists: syntactically valid queries can still misinterpret user intent. To mitigate this challenge, we propose GBV-SQL, a multi-agent framework that introduces Guided Generation with SQL2Text Back-translation Validation. In particular, a dedicated validator translates generated SQL back into natural language and checks whether its logic is aligned with the original question. Beyond the method itself, we also conduct a systematic audit of benchmark quality and introduce a typology of “Gold Errors” in Text2SQL datasets. Our analysis shows that benchmark issues can coexist with strong execution accuracy and can substantially affect evaluation outcomes. On the challenging BIRD benchmark, GBV-SQL achieves 63.23% execution accuracy, a 5.8% absolute improvement over the Deepseek-v3-based MAC-SQL setting. Under manually audited benchmark corrections, GBV-SQL reaches 96.5% (dev) and 97.6% (test) on Spider, and 90.42% on repaired BIRD dev, providing a diagnostic view of model behavior under improved gold quality. Our work contributes both a practical framework for semantic validation and an empirical analysis of benchmark integrity in Text2SQL evaluation.
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
Runkai Li, Jia Xiong, Xiuyuan He, Jieru Zhao Shanghai Jiao Tong University, Jiaqi Lv, Haowen Fang Synopsys Inc., Lei Qi Southeast University, Xi Wang Southeast University
High-Level Synthesis (HLS) improves IC development productivity by enabling hardware design from C-like languages. However, strict coding constraints and design-specific optimizations limit its widespread adoption. While recent efforts employ large language models (LLMs) to assist HLS design, they often struggle with synthesizability rules and directive semantics. To this end, we introduce ChatHLS, a multi-agent HLS design framework that leverages specialized LLMs for automated debugging and directive tuning. ChatHLS incorporates an adaptive error case expansion mechanism, combined with a reasoning-to-instruction analysis method to accurately diagnose HLS errors. To optimize hardware performance, it enables QoR-aware reasoning to learn the impact of HLS directives on the quality of results (QoR). Experimental results demonstrate that ChatHLS outperforms Gemini-3-pro with a 32.6% relative improvement in debugging, while achieving significant speedups across various HLS kernels and neural network accelerators. These results underscore the potential of ChatHLS for agile hardware development.
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