In this paper, we present EasyDistill, a comprehensive toolkit designed for effective black-box and white-box knowledge distillation (KD) of large language models (LLMs). Our framework offers versatile functionalities, including data synthesis, supervised fine-tuning, ranking optimization, and reinforcement learning techniques specifically tailored for KD scenarios. The toolkit accommodates KD functionalities for both System 1 (fast, intuitive) and System 2 (slow, analytical) models. With its modular design and user-friendly interface, EasyDistill empowers researchers and industry practitioners to seamlessly experiment with and implement state-of-the-art KD strategies for LLMs. In addition, EasyDistill provides a series of robust distilled models and KD-based industrial solutions developed by us, along with the corresponding open-sourced datasets, catering to a variety of use cases. Furthermore, we describe the seamless integration of EasyDistill into Alibaba Cloud's Platform for AI (PAI). Overall, the EasyDistill toolkit makes advanced KD techniques for LLMs more accessible and impactful within the NLP community.

Large Language Models (LLMs) have extended their impact beyond Natural Language Processing, substantially fostering the development of interdisciplinary research. Recently, various LLM-based agents have been developed to assist scientific discovery progress across multiple aspects and domains. Among these, computer-using agents, capable of interacting with operating systems as humans do, are paving the way to automated scientific problem-solving and addressing routines in researchers' workflows. Recognizing the transformative potential of these agents, we introduce ScienceBoard, which encompasses two complementary contributions: (i) a realistic, multi-domain environment featuring dynamic and visually rich scientific workflows with integrated professional software, where agents can autonomously interact via different interfaces to accelerate complex research tasks and experiments; and (ii) a challenging benchmark of 169 high-quality, rigorously validated real-world tasks curated by humans, spanning scientific-discovery workflows in domains such as biochemistry, astronomy, and geoinformatics. Extensive evaluations of agents with state-of-the-art backbones (e.g., GPT-4o, Claude 3.7, UI-TARS) show that, despite some promising results, they still fall short of reliably assisting scientists in complex workflows, achieving only a 15% overall success rate. In-depth analysis further provides valuable insights for addressing current agent limitations and more effective design principles, paving the way to build more capable agents for scientific discovery. Our code, environment, and benchmark are at https://qiushisun.github.io/ScienceBoard-Home/.

Agents based on large language models leverage tools to modify environments, revolutionizing how AI interacts with the physical world. Unlike traditional NLP tasks that rely solely on historical dialogue for responses, these agents must consider more complex factors, such as inter-tool relationships, environmental feedback and previous decisions, when making choices. Current research typically evaluates agents via multi-turn dialogues. However, it overlooks the influence of these critical factors on agent behavior. To bridge this gap, we present an open-source and high-quality benchmark $C^3$-Bench. This benchmark integrates attack concepts and applies univariate analysis to pinpoint key elements affecting agent robustness. In concrete, we design three challenges: navigate complex tool relationships, handle critical hidden information and manage dynamic decision paths. Complementing these challenges, we introduce fine-grained metrics, innovative data collection algorithms and reproducible evaluation methods. Extensive experiments are conducted on 49 mainstream agents, encompassing general fast-thinking, slow-thinking and domain-specific models. We observe that agents have significant shortcomings in handling tool dependencies, long context information dependencies and frequent policy-type switching. In essence, $C^3$-Bench aims to expose model vulnerabilities through these challenges and drive research into the interpretability of agent performance. The benchmark is publicly available at https://github.com/TencentHunyuan/C3-Benchmark.

Learning robust representations from data often requires scale, which has led to the success of recent zero-shot models such as CLIP. However, the obtained robustness can easily be deteriorated when these models are fine-tuned on other downstream tasks (e.g., of smaller scales). Previous works often interpret this phenomenon in the context of domain shift, developing fine-tuning methods that aim to preserve the original domain as much as possible. However, in a different context, fine-tuned models with limited data are also prone to learning features that are spurious to humans, such as background or texture. In this paper, we propose StarFT (Spurious Textual Alignment Regularization), a novel framework for fine-tuning zero-shot models to enhance robustness by preventing them from learning spuriosity. We introduce a regularization that aligns the output distribution for spuriosity-injected labels with the original zero-shot model, ensuring that the model is not induced to extract irrelevant features further from these descriptions. We leverage recent language models to get such spuriosity-injected labels by generating alternative textual descriptions that highlight potentially confounding features. Extensive experiments validate the robust generalization of StarFT and its emerging properties: zero-shot group robustness and improved zero-shot classification. Notably, StarFT boosts both worst-group and average accuracy by 14.30% and 3.02%, respectively, in the Waterbirds group shift scenario, where other robust fine-tuning baselines show even degraded performance.

AI-based peer review systems tend to produce shallow and overpraising suggestions compared to human feedback. Here, we evaluate how well a reasoning LLM trained with multi-objective reinforcement learning (REMOR) can overcome these limitations. We start by designing a multi-aspect reward function that aligns with human evaluation of reviews. The aspects are related to the review itself (e.g., criticisms, novelty) and the relationship between the review and the manuscript (i.e., relevance). First, we perform supervised fine-tuning of DeepSeek-R1-Distill-Qwen-7B using LoRA on PeerRT, a new dataset of high-quality top AI conference reviews enriched with reasoning traces. We then apply Group Relative Policy Optimization (GRPO) to train two models: REMOR-H (with the human-aligned reward) and REMOR-U (with a uniform reward). Interestingly, the human-aligned reward penalizes aspects typically associated with strong reviews, leading REMOR-U to produce qualitatively more substantive feedback. Our results show that REMOR-U and REMOR-H achieve more than twice the average rewards of human reviews, non-reasoning state-of-the-art agentic multi-modal AI review systems, and general commercial LLM baselines. We found that while the best AI and human reviews are comparable in quality, REMOR avoids the long tail of low-quality human reviews. We discuss how reasoning is key to achieving these improvements and release the Human-aligned Peer Review Reward (HPRR) function, the Peer Review Reasoning-enriched Traces (PeerRT) dataset, and the REMOR models, which we believe can help spur progress in the area.

Object rearrangement is a key task for household robots requiring personalization without explicit instructions, meaningful object placement in environments occupied with objects, and generalization to unseen objects and new environments. To facilitate research addressing these challenges, we introduce PARSEC, an object rearrangement benchmark for learning user organizational preferences from observed scene context to place objects in a partially arranged environment. PARSEC is built upon a novel dataset of 110K rearrangement examples crowdsourced from 72 users, featuring 93 object categories and 15 environments. To better align with real-world organizational habits, we propose ContextSortLM, an LLM-based personalized rearrangement model that handles flexible user preferences by explicitly accounting for objects with multiple valid placement locations when placing items in partially arranged environments. We evaluate ContextSortLM and existing personalized rearrangement approaches on the PARSEC benchmark and complement these findings with a crowdsourced evaluation of 108 online raters ranking model predictions based on alignment with user preferences. Our results indicate that personalized rearrangement models leveraging multiple scene context sources perform better than models relying on a single context source. Moreover, ContextSortLM outperforms other models in placing objects to replicate the target user's arrangement and ranks among the top two in all three environment categories, as rated by online evaluators. Importantly, our evaluation highlights challenges associated with modeling environment semantics across different environment categories and provides recommendations for future work.

We observe a novel phenomenon, contextual entrainment, across a wide range of language models (LMs) and prompt settings, providing a new mechanistic perspective on how LMs become distracted by ``irrelevant'' contextual information in the input prompt. Specifically, LMs assign significantly higher logits (or probabilities) to any tokens that have previously appeared in the context prompt, even for random tokens. This suggests that contextual entrainment is a mechanistic phenomenon, occurring independently of the relevance or semantic relation of the tokens to the question or the rest of the sentence. We find statistically significant evidence that the magnitude of contextual entrainment is influenced by semantic factors. Counterfactual prompts have a greater effect compared to factual ones, suggesting that while contextual entrainment is a mechanistic phenomenon, it is modulated by semantic factors. We hypothesise that there is a circuit of attention heads -- the entrainment heads -- that corresponds to the contextual entrainment phenomenon. Using a novel entrainment head discovery method based on differentiable masking, we identify these heads across various settings. When we ``turn off'' these heads, i.e., set their outputs to zero, the effect of contextual entrainment is significantly attenuated, causing the model to generate output that capitulates to what it would produce if no distracting context were provided. Our discovery of contextual entrainment, along with our investigation into LM distraction via the entrainment heads, marks a key step towards the mechanistic analysis and mitigation of the distraction problem.

Despite AI's impressive achievements, including recent advances in generative and large language models, there remains a significant gap in the ability of AI systems to handle uncertainty and generalize beyond their training data. AI models consistently fail to make robust enough predictions when facing unfamiliar or adversarial data. Traditional machine learning approaches struggle to address this issue, due to an overemphasis on data fitting, while current uncertainty quantification approaches suffer from serious limitations. This position paper posits a paradigm shift towards epistemic artificial intelligence, emphasizing the need for models to learn from what they know while at the same time acknowledging their ignorance, using the mathematics of second-order uncertainty measures. This approach, which leverages the expressive power of such measures to efficiently manage uncertainty, offers an effective way to improve the resilience and robustness of AI systems, allowing them to better handle unpredictable real-world environments.

Despite the remarkable performance of Large Language Models (LLMs), they remain vulnerable to jailbreak attacks, which can compromise their safety mechanisms. Existing studies often rely on brute-force optimization or manual design, failing to uncover potential risks in real-world scenarios. To address this, we propose a novel jailbreak attack framework, ICRT, inspired by heuristics and biases in human cognition. Leveraging the simplicity effect, we employ cognitive decomposition to reduce the complexity of malicious prompts. Simultaneously, relevance bias is utilized to reorganize prompts, enhancing semantic alignment and inducing harmful outputs effectively. Furthermore, we introduce a ranking-based harmfulness evaluation metric that surpasses the traditional binary success-or-failure paradigm by employing ranking aggregation methods such as Elo, HodgeRank, and Rank Centrality to comprehensively quantify the harmfulness of generated content. Experimental results show that our approach consistently bypasses mainstream LLMs' safety mechanisms and generates high-risk content, providing insights into jailbreak attack risks and contributing to stronger defense strategies.

Similar to LLMs, the development of vision language models is mainly driven by English datasets and models trained in English and Chinese language, whereas support for other languages, even those considered high-resource languages such as German, remains significantly weaker. In this work we present an analysis of open-weight VLMs on factual knowledge in the German and English language. We disentangle the image-related aspects from the textual ones by analyzing accu-racy with jury-as-a-judge in both prompt languages and images from German and international contexts. We found that for celebrities and sights, VLMs struggle because they are lacking visual cognition of German image contents. For animals and plants, the tested models can often correctly identify the image contents ac-cording to the scientific name or English common name but fail in German lan-guage. Cars and supermarket products were identified equally well in English and German images across both prompt languages.