The task of issue resolving aims to modify a codebase to generate a patch that addresses a given issue. However, most existing benchmarks focus almost exclusively on Python, making them insufficient for evaluating Large Language Models (LLMs) across different programming languages. To bridge this gap, we introduce a multilingual issue-resolving benchmark, called Multi-SWE-bench, covering 8 widely used programming languages: Python, Java, TypeScript, JavaScript, Go, Rust, C, and C++. In particular, this benchmark includes a total of 2,132 highquality instances, carefully curated by 68 expert annotators, ensuring a reliable and accurate evaluation of LLMs on the issue-resolving task. Based on humanannotated results, the issues are further classified into three difficulty levels. We evaluate a series of state-of-the-art models on Multi-SWE-bench, utilizing both procedural and agent-based frameworks for issue resolving. Experimental results based on Multi-SWE-bench reveal three key findings: (1) Limited generalization across languages: While existing LLMs perform well on Python issues, their ability to generalize across other languages remains limited; (2) Performance aligned with human-annotated difficulty: LLM-based agents' performance closely aligns with human-assigned difficulty, with resolved rates notably decreasing as issue complexity rises; and (3) Performance drop on cross-file issues: The performance of current methods significantly deteriorates when handling cross-file issues. These findings highlight the limitations of current LLMs and underscore the need for more robust models capable of handling a broader range of programming languages and complex issue scenarios.

Aligning large language models (LLMs) with human preferences typically demands vast amounts of meticulously curated data, which is both expensive and prone to labeling noise. We propose Stackelberg Game Preference Optimization (SGPO), a robust alignment framework that models alignment as a two-player Stackelberg game between a policy (leader) and a worst-case preference distribution (follower). The proposed SGPO guarantees O(ϵ)-bounded regret within an ϵ-Wasserstein ball, offering formal robustness to (self-)annotation noise. We instantiate SGPO with Stackelberg Self-Annotated Preference Optimization (SSAPO), which uses minimal humanlabeled "seed" preferences and iteratively self-annotates new prompts. In each iteration, SSAPO applies a distributionally robust reweighting of synthetic annotations, ensuring that noisy or biased self-labels do not derail training. Remarkably, using only 2K seed preferences--about 1/30 of standard human labels--SSAPO achieves strong win rates against GPT-4 across multiple benchmarks within three iterations.

In this paper, we introduce BMMR, a large-scale bilingual, multimodal, multidisciplinary reasoning dataset for the community to develop and evaluate large multimodal models (LMMs). BMMR comprises 110k college-level questions spanning 300 UNESCO-defined subjects, spanning diverse formats--multiplechoice, fill-in-the-blank, and open-ended QA--and sourced from both print and digital media such as books, exams, and quizzes. All data are curated and filtered via a human-in-the-loop and scalable framework, and each instance is paired with a high-quality reasoning path. The dataset is organized into two parts: BMMR-Eval that comprises 20,458high-quality instances to comprehensively assess LMMs' knowledge and reasoning across multiple disciplines in both Chinese and English; and BMMR-Train that contains 88,991 instances to support further research and development, extending the current focus on mathematical reasoning to diverse disciplines and domains. In addition, we propose the process-based multi-discipline verifier (i.e., BMMR-Verifier) for accurate and fine-grained evaluation of reasoning paths. Extensive experiments on 24 models reveal that (i) even SOTA models (e.g., o3and Gemini-2.5-Pro)

While large language models (LLMs) demonstrate impressive capabilities across numerous applications, their robustness remains a critical concern. This paper is motivated by a specific vulnerability: the order sensitivity of LLMs. This vulnerability manifests itself as the order bias observed when LLMs decide between possible options (for example, a preference for the first option) and the tendency of LLMs to provide different answers when options are reordered. The use cases for this scenario extend beyond the classical case of multiple-choice question answering to the use of LLMs for multidocument tasks and as automated evaluators in AI pipelines. We introduce Set-LLM, a novel architectural adaptation for pretrained LLMs that enables the processing of mixed set-text inputs with permutation invariance guarantees. The adaptations involve a new attention mask and new positional encodings specifically designed for sets. We provide a theoretical proof of invariance and demonstrate through experiments that Set-LLM can be trained effectively, achieving comparable or improved performance and maintaining the runtime of the original model, while altogether eliminating order sensitivity.

Despite the growing integration of retrieval-enabled AI agents into society, their safety and ethical behavior remain inadequately understood. In particular, the growing integration of LLMs and AI agents with external information sources and real-world environments raises critical questions about how they engage with and are influenced by these external data sources and interactive contexts. This study investigates how expanding retrieval access--from no external sources to Wikipedia-based retrieval and open web search--affects model reliability, bias propagation, and harmful content generation. Through extensive benchmarking of censored and uncensored LLMs and AIAgents, our findings reveal a consistent degradation in refusal rates, bias sensitivity, and harmfulness safeguards as models gain broader access to external sources, culminating in a phenomenon we term safety degradation. Notably, retrieval-enabled agents built on aligned LLMs often behave more unsafely than uncensored models without retrieval. This effect persists even under strong retrieval accuracy and prompt-based mitigation, suggesting that the mere presence of retrieved content reshapes model behavior in structurally unsafe ways. These findings underscore the need for robust mitigation strategies to ensure fairness and reliability in retrieval-enabled and increasingly autonomous AI systems. Content Warning: This paper contains examples of harmful language.

The text produced by language models (LMs) can exhibit specific'behaviors,' such as a failure to follow alignment training, that we hope to detect and react to during deployment. Identifying these behaviors can often only be done post facto, i.e., after the entire text of the output has been generated. We provide evidence that there are times when we can predict how an LM will behave early in computation, before even a single token is generated. We show that probes trained on the internal representation of input tokens alone can predict a wide range of eventual behaviors over the entire output sequence. Using methods from conformal prediction, we provide provable bounds on the estimation error of our probes, creating precise early warning systems for these behaviors.

The fundamental limitation of the behavioral cloning (BC) approach to imitation learning is that it only teaches an agent what the expert did at states the expert visited. This means that when a BC agent makes a mistake which takes them out of the support of the demonstrations, they often don't know how to recover from it. In this sense, BC is akin to giving the agent the fish - giving them dense supervision across a narrow set of states - rather than teaching them to fish: to be able to reason independently about achieving the expert's outcome even when faced with unseen situations at test-time. In response, we explore learning to search (L2S) from expert demonstrations, i.e. learning the components required to, at test time, plan to match expert outcomes, even after making a mistake. These include (1) a world model and (2) a reward model. We carefully ablate the set of algorithmic and design decisions required to combine these and other components for stable and sample/interaction-efficient learning of recovery behavior without additional human corrections. Across a dozen visual manipulation tasks from three benchmarks, our approach SAILORconsistently out-performs state-of-the-art Diffusion Policies trained via BC on the same data. Furthermore, scaling up the amount of demonstrations used for BC by 5-10 still leaves a performance gap. We find that SAILORcan identify nuanced failures and is robust to reward hacking.

Residual connection has been extensively studied and widely applied at the model architecture level. However, its potential in the more challenging data-centric approaches remains unexplored. In this work, we introduce the concept of Data Residual Matching for the first time, leveraging data-level skip connections to facilitate data generation and mitigate data information vanishing. This approach maintains a balance between newly acquired knowledge through pixel space optimization and existing core local information identification within raw data modalities, specifically for the dataset distillation task. Furthermore, by incorporating training-time refinements, our method significantly improves computational efficiency, achieving superior performance while reducing training time and peak GPU memory usage by 50%. Consequently, the proposed method Fast and Accurate Data Residual Matching for Dataset Distillation (FADRM) establishes a new stateof-the-art, demonstrating substantial improvements over existing methods across multiple dataset benchmarks in both efficiency and effectiveness. For instance, with ResNet-18 as the student model and a 0.8% compression ratio on ImageNet-1K, the method achieves 48.4% test accuracy in single-model dataset distillation and 50.9% in multi-model dataset distillation, surpassing RDED by +6.4% and outperforming

We present a novel approach to formalise and solve search-based problems using large language models, which significantly improves upon previous state-of-theart results. We demonstrate the efficacy of this approach on benchmarks like the logic puzzles tasks in ZebraLogicBench. Instead of letting the LLM attempt to directly solve the puzzles, our method prompts the model to formalise the problem in a logic-focused, human-readable, domain-specific language (DSL) called Logic.py. This formalised representation is then solved using a constraint solver, leveraging the strengths of both the language model and the solver. Our approach achieves a remarkable 65% absolute improvement over the baseline performance of Llama 3.1 70B on ZebraLogicBench, increasing its accuracy to over 90%. This significant advancement demonstrates the potential of combining language models with domain-specific languages and auxiliary tools on traditionally challenging tasks for LLMs.

The advancement of AI systems for mental health support is hindered by limited access to therapeutic conversation data, particularly for trauma treatment. We present Thousand Voices of Trauma, a synthetic benchmark dataset of 3,000 therapy conversations based on Prolonged Exposure therapy protocols for Post-traumatic Stress Disorder (PTSD). The dataset comprises 500 unique cases, each explored through six conversational perspectives that mirror the progression of therapy from initial anxiety to peak distress to emotional processing. We incorporated diverse demographic profiles (ages 18-80, M=49.3, 49.4% male, 44.4% female, 6.2% nonbinary), 20 trauma types, and 10 trauma-related behaviors using deterministic and probabilistic generation methods. Analysis reveals realistic distributions of trauma types (witnessing violence 10.6%, bullying 10.2%) and symptoms (nightmares 23.4%, substance abuse 20.8%).