Wen Wang, Yang Cao, Jing Zhang, Fengxiang He, Zheng-Jun Zha, Yonggang Wen, Dacheng Tao
Detection transformers have recently shown promising object detection results and attracted increasing attention. However, how to develop effective domain adaptation techniques to improve its cross-domain performance remains unexplored and unclear. In this paper, we delve into this topic and empirically find that direct feature distribution alignment on the CNN backbone only brings limited improvements, as it does not guarantee domain-invariant sequence features in the transformer for prediction. To address this issue, we propose a novel Sequence Feature Alignment (SFA) method that is specially designed for the adaptation of detection transformers. Technically, SFA consists of a domain query-based feature alignment (DQFA) module and a token-wise feature alignment (TDA) module. In DQFA, a novel domain query is used to aggregate and align global context from the token sequence of both domains. DQFA reduces the domain discrepancy in global feature representations and object relations when deploying in the transformer encoder and decoder, respectively. Meanwhile, TDA aligns token features in the sequence from both domains, which reduces the domain gaps in local and instance-level feature representations in the transformer encoder and decoder, respectively. Besides, a novel bipartite matching consistency loss is proposed to enhance the feature discriminability for robust object detection. Experiments on three challenging benchmarks show that SFA outperforms state-of-the-art domain adaptive object detection methods. Code has been made available at: https://github.com/encounter1997/SFA.
Reinforcement learning (RL) post-training has increasingly demonstrated strong ability to elicit reasoning behaviors in large language models (LLMs). For training efficiency, rollouts are typically generated in an off-policy manner using an older sampling policy and then used to update the current target policy. To correct the resulting discrepancy between the sampling and target policies, most existing RL objectives rely on a token-level importance sampling ratio. However, we observe unstable training dynamics in existing importance-weighted objectives under large policy lag. In this paper, we revisit LLM policy optimization under off-policy conditions and show that the exact correction for the prefix-marginal policy gradient is the cumulative prefix importance ratio. Token-level weighting does not explicitly correct for this prefix distribution shift. To stabilize LLM optimization under large off-policy drift, we propose a simple yet effective objective, Minimum Prefix Ratio (MinPRO). MinPRO replaces the unstable cumulative prefix ratio with a biased, non-cumulative surrogate based on the minimum token-level ratio observed in the preceding prefix. Extensive experiments on both dense and mixture-of-experts LLMs, across multiple mathematical reasoning benchmarks, demonstrate that MinPRO substantially improves training stability and peak performance in off-policy regimes.
Consistent submodular maximization aims to maintain a high-utility subset of size at most $k$ as new elements arrive one by one, while changing only a limited number of elements per update. Previously, [DFL+25](STOC '25) established an information-theoretic bound of $2/3$ and gave the first polynomial-time randomized algorithm with a $0.51$-approximation guarantee. However, a fundamental question remained open: Is it possible to design an efficient algorithm for consistent submodular maximization that matches the classical offline $(1-1/e)$-approximation? In this paper, we answer this question in the negative by proving that the sharp approximation threshold achievable with polynomially many value queries and worst-case constant recourse is \[β:=2-\sqrt{2}\approx0.5858<1-1/e.\] Specifically, we develop an improved poly-time randomized algorithm, which can attain a $(β-\varepsilon)$-approximation in expectation at each fixed time with $O(\varepsilon^{-2})$ worst-case recourse. The cornerstone of this algorithm is a novel anchored sampling scheme, which uses only value queries on subsets of the elements seen so far to construct a solution that remains competitive under unknown future arrivals. Conversely, we establish a matching lower bound by showing that any fixed improvement over $β$ requires either exponentially many value queries or $Ω(k)$ changes per update. Furthermore, we characterize the impact of curvature by establishing a sharp threshold of $1-(\sqrt{2}-1)\vartheta$ for consistent submodular maximization, where $\vartheta\in[0,1]$ denotes the total curvature. Surprisingly, for weighted coverage objectives, we devise a poly-time randomized algorithm that can match the optimal $(1-1/e-\varepsilon)$-approximation with $O(\varepsilon^{-1})$ worst-case recourse, thereby separating coverage maximization from general submodular maximization in the consistent settings.
Haussler's covering lemma (JCTA 1995) gives the sharp $(C/δ)^d$ metric-entropy bound for binary VC classes under normalized $\ell_1$ distance. We prove a sharp global dynamic analogue for bounded real-valued classes in normalized $\ell_p$, for every $1\le p<\infty$. Writing $d=\operatorname{sfat}_{\varepsilon/32}(\mathcal F)$, we obtain a global dynamic cover of size at most $(C/\varepsilon^p)^d$ for every horizon and $0<\varepsilon<1/2$. Here a single family of deterministic causal experts, fixed independently of the instance sequence, covers every target on every sequence. For $0<\varepsilon<1/4$, a matching worst-case lower bound $(c/\varepsilon)^{pd}$ holds already for binary classes on one fixed sequence, even with real-valued experts. The constants are universal, independent of $p$. For $p=2$, this resolves the horizon-independent sequential $\ell_2$ covering problem of Rakhlin, Sridharan, and Tewari (PTRF 2015) in a stronger global form. For $p=1$, it gives the sharp $(C/δ)^{\operatorname{Ldim}(\mathcal H)}$ global dynamic covering bound for binary classes, resolving questions of Alon, Ben-Eliezer, Dagan, Moran, Naor, and Yogev (STOC 2021) and Malliaris and Moran (Trans. AMS 2024). Our proof combines rank-decreasing correction schedules and accurate sampled prefixes with exact continuations after target identification. Shared-parent routing distributes these continuations among eligible prefixes, eliminating the logarithmic overhead.
Tiange Li, Yinan Li, Youming Qiao, Dacheng Tao, Yingjie Wang
In Grochow and Qiao (\emph{SIAM J. Comput.}, 2023), the complexity class Tensor Isomorphism (TI) was introduced and isomorphism problems for groups, algebras, and polynomials were shown to be TI-complete. In this paper, we study average-case algorithms for several TI-complete problems over finite fields, including algebra isomorphism, matrix code conjugacy, and $4$-tensor isomorphism. Our main results reveal a rich landscape of average-case complexities of Tensor Isomorphism problems, despite their equivalence in the worst-case setting. Over the finite field of order $q$, we devise (1) average-case \emph{polynomial-time} algorithms for algebra isomorphism and matrix code conjugacy that succeed for $1/Θ(q)$ fraction of inputs; (2) an average-case \emph{polynomial-time} algorithm for the $4$-tensor isomorphism that succeeds for $1/q^{Θ(1)}$ fraction of inputs; and (3) an average-case \emph{polynomial-time} algorithm for the orthogonal $3$-tensor isomorphism that succeed for $1-1/\poly(n)$ fraction of $n\times n\times n$ tensors. Our main technical contribution is to introduce the spectral properties of random matrices into algorithms for TI-complete problems. This leads to not only new algorithms but also new questions in random matrix theory over finite fields. To settle these questions, we need to extend both the generating function approach as in Neumann and Praeger (\emph{J. London Math. Soc.}, 1998) and the character sum method of Gorodetsky and Rodgers (\emph{Trans. Amer. Math. Soc.}, 2021).
Geyuan Wang, Shunyu Liu, Siyuan Liang, Chenyang Lyu, Dacheng Tao
Large Language Models (LLMs) have shown promise for empowering search agents to tackle complex information-seeking tasks. However, since acquiring external information necessitates web search, these agents are highly susceptible to various adversarial attacks (e.g., malicious information injection or stealthy context manipulation). Existing attack strategies inevitably rely on static and human-crafted templates, failing to emulate the dynamic and complex nature of real-world threats. In this paper, we propose an adversarial self-evolution framework that integrates bi-level optimization between the attacker's strategy evolution and the defender's adaptive mitigation, to expose the vulnerabilities of search agents. Specifically, we introduce the Contrastive Rollout Evolutionary Optimization (CREO) method to drive the directional evolution of the attack strategy population via contrastive evolutionary signals. To provide a stable environment for this continuous evolution, we further construct a controlled sandbox and design fine-grained metrics to quantify internal vulnerabilities. Extensive experiments across various multi-hop QA benchmarks and frontier LLMs demonstrate that the proposed framework yields attack efficacy superior to static baselines, underscoring the urgent need to develop more robust defense frameworks.
Yuxuan Du, Hanqing Zhao, Xinbiao Wang, Yiming Lu, Kaining Zhang, Dacheng Tao
AI scientists are expanding automated research across scientific domains, yet their ability to turn the scientific record into a foundation for autonomous discovery remains elusive. Here we uncover a fundamental barrier to this transition and demonstrate a route to overcome it in computational quantum science, where advances shape both what we can learn from complex quantum systems and what quantum computing can ultimately deliver. To identify this barrier, we introduce OmniQ, an expert-audited virtual research environment that converts published studies into executable reproduction tasks. Across 59 papers spanning classical simulation, quantum algorithms and quantum-property characterization, we systematically evaluate five agentic systems powered by large language models and six third-party agent harnesses. Our findings reveal that current AI scientists often fail to reproduce published results, exposing a fundamental mismatch between what scientific papers communicate and what autonomous agents require. By contrast, human expert intervention can turn these failures into successful reproductions, showing that the missing information can be identified and converted into reusable research experience. Guided by this observation, we develop Residual Agents (RA), a quantum-tailored self-improving AI scientist that distils externally verified workflow repairs into such experience. RA substantially improves research performance under fixed base models and limited adaptation budgets. For example, RA raises DeepSeek V4.1 Flash performance from 68% to 85%, matching Claude Code with Opus 5 at a third of its cost despite using a weaker base model. We integrate all reproduced results into an extensible platform built on the model context protocol, making verified research workflows directly accessible to AI scientists and allowing new scientific papers to be continuously incorporated.
Xiao Yang, Aishan Liu, Siyuan Liang, Zonghao Ying, Le Wang, Jiakai Wang, Xianglong Liu, Dacheng Tao
Large language model (LLM) agents increasingly operate in open-world environments where safety risks cannot be fully predefined. However, most existing guardrail mechanisms remain static, relying on fixed rules, prompts, or classifiers that generalize poorly to emerging risks. Here we present EVOLVE, a tool-level self-evolving defense framework that formulates agent security as a process of cumulative security evolution. Inspired by how human societies transform experiences into transferable protective knowledge, EVOLVE externalizes security knowledge into executable guardrail tools that are continuously inherited, independently audited, and iteratively refined through lifelong interaction. Across multiple benchmarks, backbone models, and embodied environments, EVOLVE consistently outperforms existing defenses across diverse risk scenarios with +22\% average F1 score. We further validate the framework in real-world robotic and autonomous agent systems. Our results suggest that scalable agent safety may require continuously evolving defense systems rather than static guardrails.
Liangbo Ning, Siyuan Liang, Yi Zhou, Yifeng Zhang, Qing Li, Wenqi Fan, Dacheng Tao
Reasoning-augmented recommender systems (RecSys) enhance user preference modeling by generating reasoning trajectories over historical interactions before making recommendations. Knowledge distillation provides a practical paradigm for transferring such reasoning capabilities from powerful large language models to lightweight RecSys. However, existing methods typically rely on a single fixed teacher, remain constrained by the teacher knowledge ceiling, and indiscriminately distill teacher-generated trajectories without assessing their actual utility for downstream recommendation. To address these limitations, we propose MUSE (Multi-Teacher Utility-guided Selection with Self-Exploration), an adaptive reasoning distillation framework that jointly improves reasoning generation and reasoning utilization. MUSE first constructs a diverse reasoning pool using multiple teacher LLMs and progressively augments it via student self-exploration, enabling the student to leverage the strengths of different teachers and gradually explore reasoning strategies beyond teacher imitation. We then develop Rank-Gain Utility Assessment (RGUA), which evaluates each candidate according to the normalized full-catalog rank improvement of the ground-truth item and identifies the most and least recommendation-beneficial trajectories. Based on these utility-aware trajectory pairs, we introduce dual-level contrastive learning objectives that guide the student generator to prefer beneficial reasoning while training the recommendation encoder to more effectively exploit recommendation-relevant preference evidence. Extensive experiments on three real-world datasets demonstrate the effectiveness and robustness of the proposed framework.
LLM-based search agents are widely used for information-seeking tasks, but their reliance on external tool returns introduces a critical security risk: web content retrieved during execution is untrusted, exposing agents to prompt injection and goal hijacking. Prior work on search-agent safety primarily focuses on static web-content injection, but modern agents issue follow-up queries and cross-check competing sources, so a single injected page is often diluted or rejected. We show that the channel delivering search and page observations is a fragile security boundary: beyond exposing the agent to a single poisoned page, a mediated search interface can repeatedly steer how the agent gathers evidence and forms its final answer. Under a constrained tool-intermediary threat model, appending only one controlled result per query can substantially increase attack success when the evidence is coordinated across the agent's trajectory. We study this setting with a strategy-driven long-horizon attack system and introduce Authority-Chain Hijack (ACH), an expert-refined strategy that turns isolated search-result and page-content manipulations into a coherent evidence chain across seemingly corroborating sources. ACH achieves the highest Overall ASR among all baselines, reaching 55.9% / 83.3% ASR / MaxN ASR on the full SafeSearch test split. We further introduce Trace-Guided Strategy Evolution (TGSE), which automatically improves attacker strategies from execution traces, replacing manual redesign with trace-driven refinement; its strongest single setting reaches 71.4% / 95.0% in held-out evaluation. Code is available at https://github.com/MBOSSLESS/Breadcrumbing-Search-Agents.
Hazel H. Kim, Jakub Binkowski, Ekaterina Fadeeva, Juan David Guerra, Yijie Hao, Zheng Yi Ho, Baixiang Huang, Buyun Liang 等 25 位作者
Hallucination remains one of the most critical challenges in deploying large language models (LLMs) reliably and safely. Yet the term is used inconsistently across the research community. We argue for a strict definition of hallucinations in LLMs. First, hallucinations should be confident incorrectness, kept separate from expressed uncertainty (UNK)1. UNK reveals that the model may be wrong, whereas a hallucination does not reveal it and is therefore more dangerous in practice. Conflating the two obscures their distinct behaviors and makes them hard to study. Second, hallucinations should be split by a simple re-elicitation test: does any semantically equivalent prompt recover the correct response? This test probes whether the relevant knowledge2 is accessible. If none does, the knowledge is likely absent or inaccessible, and the error type is a confabulation (CF) that likely requires external knowledge or retraining; if some does, the knowledge is accessible from the model but applied inconsistently, and the error type is a semantic illusion (SI) that simple prompt or decoding changes can mitigate3. Re-evaluating 14 recent hallucination methods under this taxonomy yields three findings. (1) Detection methods flag UNK more often than hallucinated responses (i.e., SI and CF). (2) Mitigation methods are far more effective for SI than for CF. (3) Latent signals are more informative than output-level signals for CF. An audit of recent hallucination papers shows that most conflate these categories. We recommend reporting CF, SI, and UNK separately, and introduce a practical labeling protocol that needs only a modest number of semantically equivalent prompts.
Bingchao Wang, Yongcheng Jing, Shunyu Liu, Chengyu Wang, Jun Huang, Jie Yang, Wei Liu, Dacheng Tao
Mobile GUI agents are typically fine-tuned 7B-scale vision-language models (VLMs) that generate each action as text. However, these VLMs decode every action token by token, and even agents that score candidate actions keep a large language model or VLM in the loop, making them slow and memory-hungry. To address these issues, we propose GUI-JEPA, to our knowledge the first decoder-free GUI agent policy built on a Joint-Embedding Predictive Architecture (JEPA), which selects actions in the latent space of a JEPA encoder pretrained on unlabelled GUI screenshots. Its action- and region-conditioned latent predictor turns the screen latent into one latent per candidate action, letting lightweight heads score all candidates against the instruction in a single batched pass, without decoding a token. We further show that the scorer relies on this latent pathway and that the predictor is trained for the decision itself: a randomly initialised predictor does just as well as one pretrained on screen dynamics. On AndroidControl, GUI-JEPA is 2.8-4.0x faster on the same GPU than a Qwen2.5-VL-7B agent we fine-tune with the same inputs, while reaching 86-92% of its in-distribution accuracy with 6% of its parameters. Code is available at https://github.com/bcwang-sjtu/GUI-JEPA
Bingchao Wang, Yongcheng Jing, Chengyu Wang, Jun Huang, Jie Yang, Wei Liu, Dacheng Tao
GUI grounding on high-resolution screens fails mostly on small targets, and zooming in is a common training-free remedy. However, most existing zooming pipelines do not match the zoom to the target's size, accept the zoomed answer unconditionally and never revisit a wrong answer. To address these issues, we propose AdaGround, a training-free framework with three components: (1) Token-Normalized Adaptive Zooming, which sets the magnification, field of view and zoom depth of each sample from the model's own estimate of the target size; (2) Verifier-in-the-Loop Refinement, which uses one marked-click Yes/No verifier to keep or reject each zoom, choose between zoom chains, trigger re-grounding and, with an existence query, decide when to abstain; and (3) Compositional Structure Grounding, which composes spreadsheet cells from their headers. On Qwen3-VL models, AdaGround surpasses state-of-the-art training-free zooming methods and many trained GUI models, improving accuracy over the backbone by 7.4 to 20.6 points on OSWorld-G and ScreenSpot-Pro, and its gains carry over to held-out UI-Vision. Code is available at https://github.com/bcwang-sjtu/AdaGround
He Du Fudan University, Bowen Li, Aijun Yang Shanghai Jiaotong University and Shanghai Artificial Intelligence Laboratory, Siyang He, Qipeng Guo Shanghai AI Laboratory, Kai Chen Shanghai AI Laboratory, Dacheng Tao Nanyang Technological University
Reliable verifiable data has become a key driver of capability gains in modern language models, enabling stable reinforcement learning with verifiable rewards and effective distillation that transfers competence across math, coding, and agentic tasks. Yet constructing generalizable synthetic verifiable data remains difficult due to hallucination-prone generation, and weak or trivial verification artifacts that fail to separate strong from weak solutions. Existing approaches often rely on task-specific heuristics or post-hoc filters that do not transfer across domains and lack a principled, universal evaluator of verifiability. In this work, we introduce an evolutionary, task-agnostic, strategy-guided, executably-checkable data synthesis framework that, from minimal seed supervision, jointly synthesizes problems, diverse candidate solutions, and verification artifacts, and iteratively discovers strategies via a consistency-based evaluator that enforces agreement between human-annotated and strategy-induced checks. This pipeline upgrades filtering into principled synthesis: it reliably assembles coherent, verifiable training instances and generalizes without domain-specific rules. Our experiments demonstrate the effectiveness of the proposed approach under both RLVR and model distillation training paradigms. The results show that training with our synthesized data yields significant improvements on both the LiveCodeBench and AgentBench-OS tasks, highlighting the robust generalization of our framework.
尚无 AI 解读，展示原文摘要
cs.CL · cs.AI
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
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
Qingyu Lu, Liang Ding Alibaba Group, Kanjian Zhang Schools of Automation, Southeast University, Jinxia Zhang Southeast University, Dacheng Tao Nanyang Technological University
The pursuit of real-time agentic interaction has driven interest in Diffusion-based Large Language Models (dLLMs) as alternatives to auto-regressive backbones, promising to break the sequential latency bottleneck. However, does such efficiency gains translate into effective agentic behavior? In this work, we present a comprehensive evaluation of dLLMs (e.g., LLaDA, Dream) across two distinct agentic paradigms: Embodied Agents (requiring long-horizon planning) and Tool-Calling Agents (requiring precise formatting).Contrary to the efficiency hype, our results on Agentboard and BFCL reveal a “bitter lesson”: current dLLMs fail to serve as reliable agentic backbones, frequently leading to systematically failure. (1) In Embodied settings, dLLMs suffer repeated attempts, failing to branch under temporal feedback. (2) In Tool-Calling settings, dLLMs fail to maintain symbolic precision (e.g. strict JSON schemas) under diffusion noise. To assess the potential of dLLMs in agentic workflows, we introduce DiffuAgent, a multi-agent evaluation framework that integrates dLLMs as plug-and-play cognitive cores. Our analysis shows that dLLMs are effective in non-causal roles (e.g., memory summarization and tool selection) but require the incorporation of causal, precise, and logically grounded reasoning mechanisms into the denoising process to be viable for agentic tasks.
尚无 AI 解读，展示原文摘要
cs.CL · cs.AI · cs.MA
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Jize Wang, Han Wu, Zhiyuan You, Yiming Song Shanghai Jiaotong University, Yijun Wang, Zifei Shan WeChat, Tencent, Yining Li Shanghai AI Laboratory, Songyang Zhang Shanghai Artificial Intelligence Laboratory 等 12 位作者
Mixture-of-Agents (MoA) improves LLM performance through layered collaboration, but its dense topology raises costs and latency. Existing methods employ LLM judges to filter responses, yet still require all models to perform inference before judging, failing to cut costs effectively. They also lack model selection criteria and struggle with large model pools, where full inference is costly and can exceed context limits. To address this, we propose RouteMoA, an efficient mixture-of-agents framework with dynamic routing. It employs a lightweight scorer to perform initial screening by predicting coarse-grained performance from the query, narrowing candidates to a high-potential subset without inference. A mixture of judges then refines these scores through lightweight self- and cross-assessment based on existing model outputs, providing posterior correction without additional inference. Finally, a model ranking mechanism selects models by balancing performance, cost, and latency. RouteMoA outperforms MoA across varying tasks and model pool sizes, reducing cost by 89.8% and latency by 63.6% in the large-scale model pool. Code is available at https://github.com/Jize-W/RouteMoA.
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
Huazheng Wang, Yongcheng Jing, Haifeng Sun, Yingjie Wang, Jingyu Wang, Jianxin Liao, Dacheng Tao Nanyang Technological University
In this paper, we investigate knowledge forgetting in large language models with a focus on its generalisation—ensuring that models forget not only specific training samples but also related implicit knowledge. To this end, we begin by identifying a broader unlearning scope that includes both target data and logically associated samples, including rephrased, subject-replaced, relation-reversed, and one-hop reasoned data. We then conduct a rigorous evaluation of 15 state-of-the-art methods across three datasets, revealing that unlearned models still recall paraphrased answers and retain target facts in their intermediate layers. This motivates us to take a preliminary step toward more generalised implicit knowledge forgetting by proposing PERMU—a novel probability perturbation-based unlearning paradigm. PERMU simulates adversarial unlearning samples to eliminate fact-related tokens from the logit distribution, collectively reducing the probabilities of all answer-associated tokens. Experiments are conducted on a diverse range of datasets, including TOFU, Harry Potter, ZsRE, WMDP, and MUSE, using models ranging from 1.3B to 13B in scale. The results demonstrate that PERMU delivers up to a 50.40% improvement in unlearning vanilla target data while maintaining a 40.73% boost in forgetting implicit knowledge. Our code can be found in the supplementary material.
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
Weixiao Zhan, Yongcheng Jing, Leszek Rutkowski AGH University of Science and Technology, Systems Research Institute Polish Academy of Sciences and University of Social Sciences, Dacheng Tao Nanyang Technological University
Knowledge distillation (KD) transfers capabilities from large language models (LLMs) to smaller students, yet it can fail unpredictably and also underpins model leakage risks. Our analysis revealed several distillation traps: tail noise, off-policy instability, and, most fundamentally, the teacher–student gap, that distort training signals. These traps manifest as overconfident hallucinations, self-correction collapse, and local decoding degradation, causing distillation to fail. Motivated by these findings, we propose a post-hoc calibration method that, to the best of our knowledge, for the first time enables control over a teacher’s distillability via reinforcement fine-tuning (RFT). Our objective combines task utility, KL anchor, and across-tokenizer calibration reward. This makes distillability a practical safety lever for foundation models, connecting robust teacher–student transfer with deployment-aware model protection. Experiments across math, knowledge QA, and instruction-following tasks show that students distilled from distillable calibrated teachers outperform SFT and KD baselines, while undistillable calibrated teachers retain their task performance but cause distilled students to collapse, offering a practical knob for both better KD and model IP protection.
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