AI 速览：作者独立复现并实证研究了直接纳什优化（DNO）的实用变体 DNO-Prct，在两个开源指令模型上考察其对齐效果、成本及与理论的偏离。
Direct Nash Optimization (DNO) aligns a language model to a general preference function by approximating the Nash equilibrium of a two-player preference game. Its practical variant, DNO-Prct, iterates on-policy sampling, LLM-based rating of the samples together with a teacher's gold response, large-margin preference-pair selection and a DPO-style contrastive update. We re-implement DNO-Prct independently and study it on two open-weight instruct models, Llama 3.1 8B Instruct and Mistral 7B Instruct v0.1, with GPT-4o or a larger open-weight model of the same family as the annotator. On AlpacaEval 2.0 with GPT-4o as judge, three iterations raise the length-controlled (LC) win rate against GPT-4o from 31.7% to 45.3% for Llama and from 6.8% to 18.4% for Mistral, and the tuned 8B model scores above Llama 3.1 70B Instruct on every reported win rate. With same-family open-weight annotators, the runs retain 62% (Llama) and 82% (Mistral) of the LC gain achieved with GPT-4o as annotator, so the rating step does not require a proprietary model. In a separate ablation configuration, variants that remove the gold responses, remove the data look-back, or accept only teacher-preferred pairs score 6 to 10 LC points below the full procedure, whereas varying the number of on-policy samples per prompt (1 to 7) moves the LC win rate by at most 2.1 points, and a margin sweep (1 to 5) run in its own pipeline by at most 2.4 points. We also report the cost of the pipeline and where the practical algorithm departs from the theory it approximates. All results are single runs, and GPT-4o is both the teacher and the evaluation judge; we discuss what this implies for the reported gains.
Ahmet Tuğrul Bayrak, Fatma Nur Korkmaz, Nur Yağmur Tuncer, Mustafa Sertaç Türkel
This paper presents TurnTurk-GPT, a text-based turn-taking predictor for Turkish obtained by fine-tuning a Turkish GPT-2 model with the TurnGPT objective on synthetic and real dialogue, and evaluates it on real Turkish conversation against rule baselines, a zero shot Turkish language model and a widely used multilingual end-of-turn detector. Evaluation uses word-level decision points on unpunctuated, lower-cased text approximating the output of streaming speech recognisers, recording-disjoint and speaker-disjoint folds, nested threshold selection and paired cluster bootstrap tests. TurnTurk-GPT outperforms the multilingual detector under both recording-disjoint and speaker-disjoint evaluation. Its gain over a zero-shot Turkish language model is largest for speakers seen in training, while on unseen speakers it matches this strong baseline. Removing the suffixes of the turn-final word reduces AP by up to 44%, against 2% for a sentence-internal word, locating the Turkish end-of-turn cue in verb-final morphology. Synthetic dialogue alone yields the weakest trained model, and when a speaker pauses the zero-shot and multilingual baselines remain close to chance. The model and the evaluation protocol are publicly released.
Scaled weight decay for Muon (Muon-SW) has been reported to train language models more efficiently than constant decoupled weight decay, but such comparisons usually reuse one learning-rate (LR) and weight-decay (WD) setting for both methods. We re-run the comparison under the MAI-Base-2 protocol: LR and WD are tuned jointly through their product for each method, on a source-scaled mixture-of-experts (MoE) scaling ladder (1.9M to 15M active parameters) that holds tokens per active parameter and data repetition fixed across sizes. We find: (1) Most of the gain is a tuning gap. At the shared setting Muon-SW wins by 0.035 bits per byte (BPB), a 1.52x compute saving; after both methods are tuned, the gap is 0.0026 BPB (95% CI [0.0018, 0.0034]), a 1.03x saving, so about 90% of the advantage disappears. (2) The small advantage that remains is real: at the smallest size, all four paired seeds favour Muon-SW at the tuned settings. (3) That advantage grows with model size: across an 8x range of active parameters, the tuned gap rises from 0.0036 to 0.0051 BPB while the untuned gap shrinks from 0.035 to 0.024, so the surviving share rises from about 10% to 21%. (4) The best weight decay moves with scale: the optimal LRWD product grows for both methods, faster for Muon-SW (ratio 1.40 to 1.91), and a law fitted on the two smallest sizes overshoots the third by 1.3-1.5x, costing up to 0.002 BPB. The whole study runs on one NVIDIA T4 GPU; we release the code and all run results.
Austin M. Y. Cheung, Harsh Vardhan Gupta, Go Nam Lui, Lok Kan Chan, Ginny Y. Wong, Simon See
Memory systems for LLM agents are evaluated mostly on social dialogue, where facts accumulate and rarely change, so remembering reduces largely to finding the turn in which something was said. Financial dialogue breaks this assumption: figures are revised, stances shift without being restated, rules stated once govern later questions, and a single company is discussed under several analytical lenses. Memory systems that retrieve the passages most similar to a question, or the most recent ones, therefore return superseded values, miss rules that share no wording with the question, and cannot recover a stance spread across sessions. We present TrackMem, a per-speaker graph memory that tracks the history of figures, separates perspectives by entity and lens, and preserves each speaker's actions in chronological order. To evaluate it, we introduce FinLoCoMo, a benchmark of 15 long-horizon investment conversations with 765 questions in two splits: one following the LoCoMo categories and one targeting four failure modes. With two readers, TrackMem outperforms existing memory systems on both splits. On the failure-mode split it reaches 73.9% macro accuracy against 69.4% for the strongest baseline, with the largest gains on questions about revised figures (86.5% against 73.4%).
XinYu Piao, Myunghoon Ryu, Joongheon Kim, Jong-Kook Kim
A large language model maintains conversational content in an external storage that is outside the model and grows as the dialogue continues. This work assumes an environment where a language model has no external memory and no prior knowledge of the future question or its answer. The motivation is to research whether a dialogue can be stored in the model parameters and later answer a question about it. This work proposes self-keyed in-parameter memorization (SKIM), a closed-form method in which the model generates its own questions about each dialogue turn and writes those questions into its parameters as the keys that later retrieve the turn's content. Generating many varied questions for each turn allows a later user question to match one even when phrased differently, and its answer is recalled from the parameters alone. Evaluated on five language models, SKIM recalls most stored dialogue turns from later user questions, and it answers from the model parameters without any external storage.
Yao Huang, Zhiyuan Zeng, Yu Wang, Zhitong Zhu, Dingcheng Duan, Hao Li, Shangzhi Xu, Yifan Wang 等 17 位作者
As frontier language models grow increasingly capable and autonomous, independent third-party evaluations are critical for characterizing their systemic safety risks. While closed-source releases are routinely accompanied by comprehensive system cards, open-weight models receive less scrutiny regarding frontier risks. To address this gap, we present an independent safety evaluation of DeepSeek V4.1 Flash-1017, the latest version of V4.1 Flash that departs from V4 with a new architecture, scaled-up training, and strong agentic capabilities. Drawing on the EU General-Purpose AI Code of Practice, we evaluate the model across 34 benchmarks, including 5 in-house suites, spanning evaluation integrity, loss of control, and harmful manipulation. Conforming to the AEF-1 standard, we compare V4.1 Flash-1017 against V4.1 Flash, earlier DeepSeek V4 models, and leading closed-source systems (GPT-6 Astra, GPT-5.6 Sol, and Gemini 3.8 Flash). Overall, V4.1 Flash-1017 shows substantial safety improvements over the V4 generation and approaches or matches leading closed-source baselines on several dimensions, while still exhibiting notable behaviors that warrant continued attention. Building on our findings, we highlight three key observations and recommendations. First, because safety can shift and even regress across minor updates, developers should integrate safety evaluation into incremental releases. Second, as final-turn outputs and end-state checks alone cannot reliably detect loss-of-control behaviors, deployments require process-level observability across both reasoning traces and action logs. Third, because public benchmarks elicit strong evaluation awareness and face data contamination risks, developing private benchmarks is essential for trustworthy frontier evaluation.
We measure the emotion geometry of six LLMs spanning Google's Gemma 2 family. Using an extraction pipeline established by Sofroniew et al. (2026) on Claude Sonnet 4.5 and replicated by Hollowell (2026) on Gemma 2 27B base, we compare emotion geometry across base and instruction-tuned variants of Gemma 2 2B, 9B, and 27B. First, all six model cells exhibit the core emotion geometry. A linear map fit between two cells places held-out emotions nearest their own vectors in the target model. Across all fifteen model pairs, fit in one direction each, this failed exactly once, and the median rank was 1 in every pair. Linear maps refit on cleaned embedding-table vectors recover every held-out emotion at rank 1 in all thirty ordered pairs. Second, scale spreads the geometry's variance across more directions without weakening its tie to valence. The leading component's share of variance falls from 31.6 percent (2B) to 29.8 (9B) to 26.7 (27B) in the base models and from 31.9 (2B) to 29.2 (9B) to 27.3 (27B) in the tuned ones, while its correlation with human valence ratings stays between .71 and .78 in all six models. Third, we project each emotion vector onto the output vocabulary and ask at what depth the vector's own emotion word enters the top 10,000 of the 256,000 tokens. In the 2B and 9B models the median emotion word is there from layer 0, while in the 27B models it begins near chance, stays outside the top 10,000 for half the network, and enters at layer 23 in one step. Fourth, tuning leaves every registered analysis-layer measure within its floor. At every size, valence correlation rises and arousal correlation falls by amounts within their floors, the latter against prediction. PC2's variance share also falls, with a pooled decrease significant after correction. Mid-late layer stability exceeds its floor and traces to base 27B's final layer. Code, data, and audit reports are released.
Latency-sensitive spoken agents must know not only which tool to call, but also when they have heard enough to act. Waiting for the turn to complete delays execution even when the request is already actionable. However, neither existing spoken agents nor spoken dialogue corpora define the earliest point at which a request becomes actionable. We introduce a synthetic spoken-dialogue pipeline that provides earliest-actionable-point supervision, marking when the observed speech grounds a tool and its required arguments. Through controlled task generation, speech synthesis and alignment, and gated annotation, we construct 26,694 dialogues totaling 786 hours across 48 domains, with disjoint training and test argument values. We additionally create an audio rendering of $τ^2$-bench for external evaluation. We train an omni-modal language model on this corpus, enabling causal, chunk-level decisions and partial-transcript generation for downstream call completion. The model achieves tool-state macro-F1 scores of 0.48 and 0.47 on held-out dialogues and the external benchmark, respectively. For correct calls, the estimated mean lead over a turn-end cascade, using the same formatter, is 1.8-2.0 seconds in-domain and 1.1 seconds on $τ^2$-bench. These results demonstrate a data-driven path toward spoken agents that act while listening. We will open-source the pipeline and the training scripts.
Jin Zhao, Marta Knežević School of Computer and Communication Sciences, EPFL - EPF Lausanne, Tanja Käser EPFL
Large Language Models (LLMs) are increasingly used in education, yet their default helpfulness often conflicts with pedagogical principles. Prior work evaluates pedagogical quality via answer leakage–the disclosure of complete solutions instead of scaffolding–but typically assumes well-intentioned learners, leaving tutor robustness under student misuse largely unexplored. In this paper, we study scenarios where students behave adversarially and aim to obtain the correct answer from the tutor. We evaluate a broad set of LLM-based tutor models, including different model families, pedagogically aligned models, and a multi-agent design, under a range of adversarial student attacks. We adapt six groups of adversarial and persuasive techniques to the educational setting and use them to probe how likely a tutor is to reveal the final answer. We evaluate answer leakage robustness using different types of in-context adversarial student agents, finding that they often fail to carry out effective attacks. We therefore introduce an adversarial student agent that we fine-tune to jailbreak LLM-based tutors, which we propose as the core of a standardized benchmark for evaluating tutor robustness. Finally, we present simple but effective defense strategies that reduce answer leakage and strengthen the robustness of LLM-based tutors in adversarial scenarios.
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
Mengyu Xiang, Tinghao Chen, Boxu Han, Qiudan Li, Shu Wu Institute of automation, Chinese academy of sciences, Daniel Dajun Zeng Institute of Automation, Chinese Academy of Sciences
As social media grows, harmful information spreads rapidly across platforms and evolves over time, showing cross-platform and cross-temporal variations. Existing methods rely on fixed model parameters during training, which fail to handle substantial semantic discrepancies, leading to Out-Of-Distribution (OOD) problems. While test-time tuning enables dynamic parameter adjustment, it may lead to excessive adaptation to individual samples. The key challenge is how to adapt to semantic variations during testing while preventing overfitting from continuous tuning. To tackle this issue, this paper proposes RLAT, a reinforcement learning (RL)–guided adaptive tuning method for harmful text detection. First, a tuning joint optimization module is designed to update parameters and adapt to semantic variations during testing. It tunes the model by optimizing consistency loss and applying word-level attention constraints to reduce over-reliance on local words and learn a more robust global representation. Then, to mitigate overfitting caused by continuous tuning, a RL–guided adaptive decision model is introduced to direct the tuning process. It reduces the influence of local samples by selecting data and controlling parameter updates, thereby improving overall test performance. Experimental results show that the RLAT outperforms state-of-the-art baselines in cross-platform and cross-temporal scenarios across multiple public 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
Large language models (LLMs) have demonstrated better safety performance in high-resource languages than in low-resource languages. We attribute this issue as a mismatch gap between language-agnostic semantic understanding ability and language dominant safety alignment biased toward high-resource languages. Based on above insights, we empirically identify the semantic bottleneck in LLMs: intermediate layers in which the geometry of model representations is governed primarily by shared semantic content rather than language identity. Then, we propose Language-Agnostic Semantic Alignment (LASA), which anchors safety alignment directly in semantic bottlenecks. Experiments show that LASA substantially improves safety across all languages: average attack success rate (ASR) drops from 24.7% to 2.8% on LLaMA-3.1-8B-Instruct and remains within 3–4% across Qwen2.5 and Qwen3 Instruct models (7B–32B). Besides, our analysis and method offer a representation-level perspective on LLM safety, suggesting that safety alignment requires anchoring safety understanding not in surface text, but in the model’s language-agnostic semantic space.
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
Zechen Li, Qiannan Zhu Beijing Normal University, Mei Wang Beijing Normal University, Jia Li Beijing Normal University, Hua Huang Beijing Normal University
Equipping Large Language Models (LLMs) with pedagogical tutoring capabilities holds significant promise for education. Existing approaches simulate tutor behaviors or preferences and use them to prompt or fine-tune LLMs for dialogue tutoring. However, such methods often fail to sustain high-quality pedagogical conversations that provide explicit stepwise scaffolding and adapt to learners’ evolving cognitive states. To address this, we propose ScaffoldLM, a planning-guided tutoring framework with an assessment-driven memory for multi-turn math dialogue tutoring. ScaffoldLM first generates a stepwise pedagogical plan from solution steps, which serves as a stable backbone for explicit scaffolding. During tutoring, the tutoring memory is updated by an assessment-driven control loop that infers the learner’s cognitive state, evaluates whether the current step target is met, and adaptively selects tutoring actions. The plan, step-level progress, inferred learner states, and dialogue history are maintained in memory to support coherent multi-turn guidance. Experiments on multi-turn math tutoring benchmarks demonstrate that ScaffoldLM substantially improves pedagogical tutoring quality over strong baselines. Code is publicly available at https://github.com/BNU-ERC-ITEA/ScaffoldLM.
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
Lijun Wu Shanghai AI Laboratory, Wentao Zhang Peking University, Conghui He Shanghai AI Lab
As the development of Large Language Models (LLMs) matures, the focus of the research community is undergoing a critical shift from a purely model-centric to a data-centric paradigm. It is now evident that the quality, diversity, and composition of training data—not merely its scale—are the primary drivers of a model’s advanced capabilities, from complex reasoning to reliable instruction following. However, acquiring and curating such high-quality data remains a significant bottleneck. This tutorial provides a comprehensive and practical guide to the state-of-the-art in data research directions for LLMs. We structure the tutorial around the two core pillars of modern data strategy: intelligent data selection and advanced data synthesis. In the first part, we delve into methods for curating the most valuable information from vast, noisy datasets, covering techniques like LLM-as-a-judge for automated quality filtering and active learning for maximizing annotation efficiency. The second part explores the synthetic data revolution, detailing paradigms that range from generating complex reasoning traces (e.g., Chain-of-Thought) to deploying sophisticated multi-agent workflows that can autonomously create high-quality, diverse instruction data from raw seeds. Finally, we will conclude with a practical overview of open-source tools and platforms that facilitate these data-centric workflows, empowering researchers and practitioners to build better models through better data. Attendees will leave with a principled framework and actionable insights for designing and implementing the advanced data strategies required to build the next generation of powerful, specialized, and aligned LLMs.
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 5: Tutorial Abstracts) · 2026 · 正式发表
Akhil Arora Aarhus University, Vishrav Chaudhary Meta, Julia Kreutzer Cohere Labs, Nearchos Potamitis Aarhus University, Nouha Dziri Allen Institute for AI, Niket Tandon Microsoft Research
As large language models (LLMs) increasingly tackle reasoning-heavy tasks, from mathematics to commonsense to multilingual understanding, researchers face three pressing questions: How well do models reason? How can we make them reason better? What are the next frontiers in LLM reasoning? This tutorial answers these questions through a unified view of LLM reasoning. This tutorial explores comprehensive evaluation strategies to assess the reasoning abilities of models and discusses two types of methods to improve models’ reasoning: advanced inference time methods, such as structured and self-improvement inference methods, and (ii) post-training methods, such as RLHF, DPO, and GRPO that aim to make LLMs think more like humans. The tutorial explores these technical discussions while maintaining a practical outlook through illustrative demos and short guided hands-on exercises. The tutorial is designed for both researchers and practitioners seeking practical insights into LLM reasoning.
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 5: Tutorial Abstracts) · 2026 · 正式发表
Joanne Boisson Cardiff University, Jose Camacho-Collados Cardiff University, Elisa Sanchez-Bayona University of the Basque Country, Rodrigo Agerri University of the Basque Country
This tutorial will be of the type Introductory to CL/NLP topics, providing an overview of the metaphor processing field. The tutorial will discuss the influence of various metaphor theories on the creation of annotated resources and models. We will particularly focus on recent directions opened by LLMs for metaphor interpretation in multilingual and multimodal settings. Two types of audience may benefit from this tutorial: researchers in the humanities and computational social sciences interested in automatic or semi automatic metaphor analysis, and NLP researchers interested in understanding metaphor processing and improving metaphor modeling.
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 5: Tutorial Abstracts) · 2026 · 正式发表
Predicting enzyme kinetic parameters quantifies how efficiently an enzyme catalyzes a specific substrate under defined biochemical conditions. Canonical parameters such as the turnover number (k_\text cat ), Michaelis constant (K_\text m ), and inhibition constant (K_\text i ) depend jointly on the enzyme sequence, the substrate chemistry, and the conformational adaptation of the active site during binding. Many learning pipelines simplify this process to a static compatibility problem between the enzyme and substrate, fusing their representations through shallow operations and regressing a single value. Such formulations overlook the staged nature of catalysis, which involves both substrate recognition and conformational adaptation. In this regard, we reformulate kinetic prediction as a staged multimodal conditional modeling problem and introduce the Enzyme-Reaction Bridging Adapter (ERBA), which injects cross-modal information via fine-tuning into Protein Language Models (PLMs) while preserving their biochemical priors. ERBA performs conditioning in two stages: Molecular Recognition Cross-Attention (MRCA) first injects substrate information into the enzyme representation to capture specificity; Geometry-aware Mixture-of-Experts (G-MoE) then integrates active-site structure and routes samples to pocket-specialized experts to reflect induced fit. To maintain semantic fidelity, Enzyme-Substrate Distribution Alignment (ESDA) enforces distributional consistency within the PLM manifold in a reproducing kernel Hilbert space. Experiments across three kinetic endpoints and multiple PLM backbones, ERBA delivers consistent gains and stronger out-of-distribution performance compared with sequence-only and shallow-fusion baselines, offering a biologically grounded route to scalable kinetic prediction and a foundation for adding cofactors, mutations, and time-resolved structural cues.
尚无 AI 解读，展示原文摘要
cs.CV · cs.CL
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2026 · 正式发表
Humanoid robots have achieved significant progress in motion generation and control, exhibiting movements that appear increasingly natural and human-like. Inspired by the Turing Test, we propose the Motion Turing Test, a framework that evaluates whether human observers can discriminate between humanoid robot and human poses using only kinematic information. To facilitate this evaluation, we present the Human-Humanoid Motion (HHMotion) dataset, which consists of 1,000 motion sequences spanning 15 action categories, performed by 11 humanoid models and 10 human subjects. All motion sequences are converted into SMPL-X representations to eliminate the influence of visual appearance. We recruited 30 annotators to rate the human-likeness of each pose on a 0-5 scale, resulting in over 500 hours of annotation. Analysis of the collected data reveals that humanoid motions still exhibit noticeable deviations from human movements, particularly in dynamic actions such as jumping, boxing, and running. Building on HHMotion, we formulate a human-likeness evaluation task that aims to automatically predict human-likeness scores from motion data. Despite recent progress in multimodal large language models, we find that they remain inadequate for assessing motion human-likeness. To address this, we propose a simple baseline model and demonstrate that it outperforms several contemporary LLM-based methods. The dataset, code, and benchmark release publicly at http://www.lidarhumanmotion.net/mtt/
尚无 AI 解读，展示原文摘要
cs.CV · cs.CL
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2026 · 正式发表