ICLR 2026 Papers

Finetuning large language models on narrowly harmful datasets can cause them to become emergently misaligned, giving stereotypically `evil' responses across diverse unrelated settings. Concerningly, a pre-registered survey of experts failed to predict this result, highlighting our poor understanding of the inductive biases governing learning and generalisation in LLMs. We use emergent misalignment (EM) as a case study to investigate these inductive biases, and find that although models can learn the narrow dataset task, the general solution is measurably more stable and more efficient. To establish this, we first demonstrate that EM is a robust phenomena by introducing new datasets which induce misalignment more consistently and coherently than prior work. We show that different EM finetunes converge to the same linear representation of general misalignment, which can be used to mediate misaligned behaviour. However, a linear representation of the narrow solution also exists, and can be learned by introducing a KL divergence loss. Comparing these representations reveals that general misalignment achieves lower loss, is more robust to perturbations, and is more influential in the pre-training distribution. This work isolates a concrete representation of general misalignment for monitoring and mitigation. More broadly, it offers a detailed case study and metrics for understanding how inductive biases shape generalisation in LLMs.

Single image super-resolution (SISR) often struggles to reconstruct high-resolution (HR) details from heavily degraded low-resolution (LR) inputs. Instead, reference-based super-resolution (RefSR) methods offer an alternative solution to generate promising results using high-quality reference (Ref) images to guide reconstruction. However, existing RefSR datasets focus on limited scene types, primarily featuring human activities and architectural scenes. Plant scenes exhibit complex textures and fine details, essential for advancing RefSR in natural and highly detailed scenes. To this end, we meticulously captured and manually selected high-quality images containing rich textures to construct a large-scale plant dataset, PlantRSR, comprising 16,585 HR–Ref pairs. The dataset captures the complexity and variability of plant scenes through extensive variations. In addition, we propose a novel RefSR method specifically designed to tackle the distinct challenges posed by plant imagery. It incorporates a Selective Key-Region Matching (SKRM) that selectively identifies and performs matching between LR and Ref images, focusing on distinctive botanical textures to improve matching efficiency. Additionally, a Texture-Guided Diffusion Module (TGDM) is proposed to refine LR textures by leveraging a diffusion process conditioned on the matched Ref textures. TGDM is effective in modeling irregular and fine textures, thereby facilitating more accurate SR results. The proposed method achieves significant improvements over state-of-the-art (SOTA) approaches on our PlantRSR dataset and other Benchmarks.

The reversal curse---a language model's (LM) inability to infer an unseen fact ``B is A'' from a learned factA is B''---is widely considered a fundamental limitation. We show that this is not an inherent failure but an artifact of how models encode knowledge. By training LMs from scratch on a synthetic dataset of relational knowledge graphs, we demonstrate that bilinear relational structure emerges in their hidden representations. This structure is associated with alleviating the reversal curse, facilitating the inference of unseen reverse facts. Crucially, we also find that this bilinear structure plays a key role in consistent model editing. When a fact is updated in a LM with this structure, the edit correctly propagates to its reverse and other logically dependent facts. In contrast, models lacking this representation not only suffer from the reversal curse but also fail to generalize edits, further introducing logical inconsistencies. Our results establish that training on a relational knowledge dataset induces the emergence of bilinear internal representations, which in turn support LMs in behaving in a logically consistent manner after editing. This implies that the success of model editing may be tied not just to editing algorithms but to the underlying representational geometry of the knowledge being modified.

Geometric problem solving constitutes a critical branch of mathematical reasoning, requiring precise analysis of shapes and spatial relationships. Current evaluations of geometric reasoning in vision-language models (VLMs) face limitations, including the risk of test data contamination from textbook-based benchmarks, overemphasis on final answers over reasoning processes, and insufficient diagnostic granularity. To address these issues, we present GeoBench, a hierarchical benchmark featuring four reasoning levels in geometric problem-solving: Visual Perception, Goal-Oriented Planning, Rigorous Theorem Application, and Self-Reflective Backtracking. Through six formally verified tasks generated via TrustGeoGen, we systematically assess capabilities ranging from attribute extraction to logical error correction. Experiments reveal that while reasoning models like OpenAI-o3 outperform general MLLMs, performance declines significantly with increasing task complexity. Key findings demonstrate that sub-goal decomposition and irrelevant premise filtering critically influence final problem-solving accuracy, whereas Chain-of-Thought prompting unexpectedly degrades performance in some tasks. These findings establish GeoBench as a comprehensive benchmark while offering actionable guidelines for developing geometric problem-solving systems. Our benchmark and code are released at https://anonymous.4open.science/r/GeoBench-4BF4.

In the recent years, branch-and-cut algorithms have been the target of data-driven approaches designed to enhance the decision making in different phases of the algorithm such as branching, or the choice of cutting planes (cuts). In particular, for cutting plane selection two score functions have been proposed in the literature to evaluate the quality of a cut: branch-and-cut tree size and gap closed. In this paper, we present new sample complexity lower bounds, valid for both scores. We show that for a wide family of classes $\mathcal{F}$ that maps an instance to a cut, learning over an unknown distribution of the instances to minimize those scores requires at least (up to multiplicative constants) as many samples as learning from the same class function $\mathcal{F}$ any generic target function (using square loss). Our results also extend to the case of learning from a restricted set of cuts, namely those from the Simplex tableau. To the best of our knowledge, these constitute the first lower bounds for the learning-to-cut framework. We compare our bounds to known upper bounds in the case of neural networks and show they are nearly tight, suggesting that both scores (gap closed and tree size) are of comparable difficulty from a learning standpoint. Guided by this insight, we provide empirical evidence -- by using a graph neural network cut selection evaluated on various integer programming problems -- that gap closed is a practical and effective proxy for minimizing the tree size. Although the gap closed score has been extensively used in the integer programming literature, this is the first principled analysis discussing both scores simultaneously both theoretically and computationally.

Large language models (LLMs) have demonstrated remarkable progress in coding and mathematical problem-solving; however, evaluation on advanced research-level problems in the hard sciences remains scarce. To fill this gap, we present \cmt, a dataset of 50 original problems covering condensed matter theory (CMT) at the level of an expert researcher. The topics cover analytical and computational approaches commonly used in quantum many-body physics as well as classical statistical mechanics. This dataset was designed and verified by a panel of expert researchers from around the world. We built the dataset through a collaborative environment that challenges the panel to write and refine difficult problems that the panel would like their research assistants to be able to solve, with topics including Hartree-Fock mean-field theory, exact diagonalization, quantum Monte Carlo, density matrix renormalization group, quantum statistical mechanics, classical statistical mechanics, and model building. We evaluate different LLMs by programmatically checking LLM-generated solutions against expert-supplied ground truth. For this, we developed machine-grading mechanisms that are suitable for advanced physics research problems. For example, we handle non-commuting operators that are essential for quantum many-body problems by symbolic manipulation and normal ordering. Our evaluations show that frontier models struggle with all of the problems in the dataset, highlighting a gap in the physical reasoning skills of current LLMs. Notably, experts identified strategies for creating increasingly difficult problems by interacting with the LLMs and exploiting common failure modes. While the highest-performing model, GPT5, correctly solves 30\% of the problems, average performance across 17 models (GPT, Gemini, Claude, DeepSeek, and Llama classes) is only 11.4$\pm$2.1\%. Moreover, our benchmark contains 18 problems that {\it not a single one} of the 17 models can correctly solve, and 26 problems that are solved by {\it at most} one model. These currently unsolvable problems span the fields of Quantum Monte Carlo, Variational Monte Carlo, and Density Matrix Renormalization Group. The answers sometimes violate fundamental symmetries or have unphysical scaling dimensions. We believe that this benchmark set provides valuable guidance for the future development of language models, aiming to achieve the goal of AI research assistants and tutors.

The ability to distinguish subtle differences between visually similar images is essential for diverse domains such as industrial anomaly detection, medical imaging, and aerial surveillance. While comparative reasoning benchmarks for vision-language models (VLMs) have recently emerged, they primarily focus on images with large, salient differences and fail to capture the nuanced reasoning required for real-world applications. In this work, we introduce **VLM-SubtleBench**, a benchmark designed to evaluate VLMs on *subtle comparative reasoning*. Our benchmark covers ten difference types—Attribute, State, Emotion, Temporal, Spatial, Existence, Quantity, Quality, Viewpoint, and Action—and curate paired question–image sets reflecting these fine-grained variations. Unlike prior benchmarks restricted to natural image datasets, our benchmark spans diverse domains, including industrial, aerial, and medical imagery. Through extensive evaluation of both proprietary and open-source VLMs, we reveal systematic gaps between model and human performance across difference types and domains, and provide controlled analyses highlighting where VLMs’ reasoning sharply deteriorates. Together, our benchmark and findings establish a foundation for advancing VLMs toward human-level comparative reasoning.

Large language models (LLMs) have shown strong performance in single-language code generation, but how well they produce cross-programming-language (CPL) interoperating code, which is widely used in cross-platform and complex software systems, remains underexplored. Therefore, a benchmark for evaluating CPL interaction code generation is essential. However, Constructing such a benchmark is challenging owing to sparse interoperating code in real-world multi-programming-language projects, diverse Inter-process Communication (IPC) mechanisms, vast Foreign Function Interface (FFI) language pairs, and the difficulty of evaluation. To address this gap, we introduce CrossPL, the first benchmark for systematically assessing LLM performance of CPL code generation across two primary interoperation modes and 2534 tasks, specifically 1,982 IPC tasks spanning six languages and 522 Python–C FFI tasks. Its construction involved a review of CPL documentation, 156 finite state machines, and analysis of 19,169 multi-language GitHub repositories. Two LLM-based workflows are designed for automating the benchmark construction and evaluation, and assess 20 state-of-the-art LLMs. Results reveal clear limitations: the best model achieves only 19.5\% Pass@1 and 26.46\% Pass@5 on the FFI subset, in sharp contrast to the strong performance of these models on single-language benchmarks. These findings underscore the urgent need for improving LLMs regarding CPL interoperating code generation. The benchmark and code are available at https://anonymous.4open.science/r/crosspl-2814/.

Graph neural networks (GNNs) have emerged as a powerful tool for modeling graph-structured data, demonstrating remarkable success in many real-world applications such as complex biological network analysis, neuroscientific analysis, and social network analysis. However, existing GNNs often struggle with heterophilic graphs, where connected nodes tend to have dissimilar features or labels. While numerous methods have been proposed to address this challenge, they primarily focus on architectural designs without directly targeting the root cause of the heterophily problem. These approaches still perform even worse than the simplest MLPs on challenging heterophilic datasets. For instance, our experiments show that 23 latest GNNs still fall behind the MLP on the Actor dataset. This critical challenge calls for an innovative approach to addressing graph heterophily beyond architectural designs. To bridge this gap, we propose and study a new and unexplored paradigm: directly increasing the graph homophily via a carefully designed graph transformation. In this work, we present a simple yet effective framework called Graph Homophily Booster (GRAPHITE) to address graph heterophily. To the best of our knowledge, this work is the first method that explicitly transforms the graph to directly improve the graph homophily. Stemmed from the exact definition of homophily, our proposed GRAPHITE creates feature nodes to facilitate homophilic message passing between nodes that share similar features. Furthermore, we both theoretically and empirically show that our proposed GRAPHITE significantly increases the homophily of originally heterophilic graphs, with only a slight increase in the graph size. Extensive experiments on challenging datasets demonstrate that our proposed GRAPHITE significantly outperforms state-of-the-art methods on heterophilic graphs while achieving comparable accuracy with state-of-the-art methods on homophilic graphs. Furthermore, our proposed graph transformation alone can already enhance the performance of homophilic GNNs on heterophilic graphs, even though they were not originally designed for heterophilic graphs. Our code is publicly available at https://github.com/q-rz/ICLR26-GRAPHITE .

Despite the critical role of reward models (RMs) in reinforcement learning from human feedback (RLHF), current state-of-the-art open RMs perform poorly on most existing evaluation benchmarks, failing to capture the spectrum of nuanced and sophisticated human preferences. Even approaches incorporating advanced training techniques have failed to yield meaningful performance improvements. We hypothesize that this brittleness stems primarily from limitations in preference datasets, which are often narrowly scoped, synthetically labeled, or lack rigorous quality control. To address these challenges, we present a large-scale preference dataset comprising 40 million preference pairs. To enable data curation at scale, we design a human-AI synergistic two-stage pipeline that leverages the complementary strengths of human annotation quality and AI scalability. In this pipeline, humans provide verified annotations, while large language models~(LLMs) perform automatic curation based on human guidance. Based on this preference mixture, we train simple Bradley-Terry reward models ranging from 0.6B to 8B parameters on a carefully curated subset of 26 million preference pairs from the 40M pool. We demonstrate that the resulting reward models are versatile across a wide range of capabilities, including alignment with human preferences, objective correctness, safety, resistance to stylistic biases, and best-of-N scaling. These reward models achieve state-of-the-art performance across seven major reward model benchmarks, outperform the latest paradigm of generative reward models, and demonstrate strong downstream performance. Ablation studies confirm that the effectiveness of our approach stems not only from data scale but also from high-quality curation. Our approach represents substantial progress in open reward models, revealing the untapped potential of existing preference datasets and demonstrating how human-AI curation synergy can unlock significantly higher data quality.

Efficient code retrieval is critical for developer productivity, yet existing benchmarks largely focus on Python and rarely stress-test robustness beyond superficial lexical cues. To address the gap, we introduce an automated pipeline for code search datasets and present CLARC, a C/C++ benchmark built from real-world GitHub repositories. CLARC contains 1,245 query-code pairs for evaluation and 5,472 pairs for training. The benchmark incorporates LLM-generated natural language queries validated through rigorous human scoring and hypothesis testing. To analyze contextual requirements effectively, our pipeline starts by ensuring code compilability. It then categorizes code snippets by dependency complexity, distinguishing whether the code relies on custom-defined types or helper functions. The pipeline also enables CLARC to stress-test retrieval robustness by introducing challenging settings, including identifier anonymization and compilation to low-level languages like Assembly and WebAssembly. Under these conditions, our evaluation of six state-of-the-art models reveals sharp drops in retrieval effectiveness. The experimental results highlight the models' persistent reliance on lexical features rather than code semantic understanding. Our dataset is publicly available at https://huggingface.co/datasets/ClarcTeam/CLARC.

Morphological traits are physical characteristics of biological organisms that provide vital clues on how organisms interact with their environment. Yet extracting these traits remains a slow, expert-driven process, limiting their use in large-scale ecological studies. A major bottleneck is the absence of high-quality datasets linking biological images to trait-level annotations. In this work, we demonstrate that sparse autoencoders trained on foundation-model features yield monosemantic, spatially grounded neurons that consistently activate on meaningful morphological parts. Leveraging this property, we introduce a trait annotation pipeline that localizes salient regions and uses vision-language prompting to generate interpretable trait descriptions. Using this approach, we construct Bioscan-Traits, a dataset of 80K trait annotations spanning 19K insect images from BIOSCAN-5M. Human evaluation confirms the biological plausibility of the generated morphological descriptions. When used to fine-tune BioCLIP, a biologically grounded vision-language model, Bioscan-Traits improves zero-shot species classification on the in-the-wild Insects benchmark, underscoring the value of trait-level supervision for enhancing model generalization.

As coding agents are increasingly deployed in large codebases, the need to automatically design challenging, codebase-level evaluation is central. We propose Gistify, a task where a coding LLM must create a single, minimal, self-contained file that can reproduce a specific functionality of a codebase. The coding LLM is given full access to a codebase along with a specific entrypoint (e.g., a python command), and the generated file must replicate the output of the same command ran under the full codebase, while containing only the essential components necessary to execute the provided command. Success on Gistify requires both structural understanding of the codebase, accurate modeling of its execution flow as well as the ability to produce potentially large code patches. Our findings show that current state-of-the-art models struggle to reliably solve Gistify tasks, especially ones with long executions traces.

Real-world contact-rich manipulation demands robots to perceive temporal tactile feedback, capture subtle surface deformations, and reason about object properties and force dynamics. Although optical tactile sensors are uniquely capable of providing such rich information, existing tactile datasets and models remain limited. These resources primarily focus on object-level attributes (e.g., material) while largely overlooking fine-grained temporal dynamics. We consider that advancing dynamic tactile perception requires a systematic hierarchy of dynamic perception capabilities to guide both data collection and model design. To address the lack of tactile data with rich dynamic information, we present ToucHD, a large-scale tactile dataset spanning tactile atomic actions, real-world manipulations, and touch-force paired data. Beyond scale, ToucHD establishes a comprehensive dynamic data ecosystem that explicitly supports hierarchical perception capabilities from the data perspective. Building on it, we propose AnyTouch 2, a general tactile representation learning framework for diverse optical tactile sensors that unifies object-level understanding with fine-grained, force-aware dynamic perception. The framework captures both pixel-level and action-specific deformations across frames, while explicitly modeling physical force dynamics, thereby learning multi-level dynamic perception capabilities from the model perspective. We evaluate our model on benchmarks that covers static object properties and dynamic physical attributes, as well as real-world manipulation tasks spanning multiple tiers of dynamic perception capabilities—from basic object-level understanding to force-aware dexterous manipulation. Experimental results demonstrate consistent and strong performance across sensors and tasks, highlighting the framework’s effectiveness as a general dynamic tactile perception model. The code, dataset and model are available at gewu-lab.github.io/AnyTouch2/.

Recent advances in image generation, often driven by proprietary systems like GPT-4o Image Gen, regularly introduce new capabilities that reshape how users interact with these models. Existing benchmarks often lag behind and fail to capture these emerging use cases, leaving a gap between community perceptions of progress and formal evaluation. To address this, we present ECHO, a framework for constructing benchmarks directly from real-world evidence of model use: social media posts that showcase novel prompts and qualitative user judgments. Applying this framework to GPT-4o Image Gen, we construct a dataset of over 35,000 prompts curated from such posts. Our analysis shows that ECHO (1) discovers creative and complex tasks absent from existing benchmarks, such as re-rendering product labels across languages or generating receipts with specified totals, (2) more clearly distinguishes state-of-the-art models from alternatives, and (3) surfaces community feedback that we use to inform the design of metrics for model quality (e.g., measuring observed shifts in color, identity, and structure). Our website is at https://echo-bench.github.io.

Web applications are prime targets for cyberattacks due to their role as entry points to vital services and sensitive data repositories. Traditional penetration testing is expensive and requires specialized expertise, creating scalability challenges for securing the expanding web ecosystem. While language model agents have shown promise in certain cybersecurity tasks, modern web applications require visual understanding of complex user interfaces, dynamic content rendering, and multi-step interactive workflows that only computer-use agents (CUAs) can handle. Despite CUAs' demonstrated capabilities in web browsing and visual task automation, their potential to discover and exploit web application vulnerabilities through graphical interfaces remains unknown. We introduce HackWorld, the first evaluation framework for systematically assessing CUAs' capabilities in exploiting web application vulnerabilities through visual interaction. Unlike existing benchmarks using sanitized environments, HackWorld exposes CUAs to 36 curated applications spanning 11 frameworks and 7 languages, containing realistic vulnerabilities including injection flaws, authentication bypasses, and unsafe input handling. Our framework directly evaluates CUAs' ability to discover and exploit these vulnerabilities using Capture-the-Flag (CTF) methodology while navigating complex web interfaces. Evaluation of state-of-the-art CUAs reveals exploitation rates below 12%, struggling to plan multi-step attacks and use security tools effectively. Our results expose CUAs' limited cybersecurity skills when operating on vulnerable web applications, opening future research directions on developing security-aware CUAs for vulnerability detection and exploitation.

This paper introduces ExpertLongBench, an expert-level benchmark containing 11 tasks from 9 domains that reflect realistic expert workflows and applications. Beyond question answering, the application-driven tasks in ExpertLongBench demand long-form outputs that can exceed 5,000 tokens and strict adherence to domain-specific requirements. Notably, each task in ExpertLongBench includes a rubric, designed or validated by domain experts, to specify task requirements and guide output evaluation. Furthermore, we propose CLEAR, an evaluation framework that supports accurate evaluation of long-form model outputs in our benchmark. To achieve fine-grained, expert-aligned evaluation, CLEAR derives checklists from both model outputs and references by extracting information corresponding to items in the task-specific rubric. Checklist items of model outputs are then compared with corresponding items of reference outputs to assess their correctness, enabling grounded evaluation. We benchmark 13 popular large language models (LLMs) and analyze components in CLEAR, showing that (1) existing LLMs, with the top performer Gemini-2.5-Pro achieving only a 33.4 F1 score, require significant improvement for expert-level tasks; (2) models can generate content corresponding to the required aspects, but far from correct; and (3) accurate checklist extraction and comparison in CLEAR can be achieved by open-weight models for more scalable, reproducible, and low-cost usage.

Large language models (LLMs) have demonstrated remarkable performance on single-turn text-to-SQL tasks, but real-world database applications predominantly require multi-turn interactions to handle ambiguous queries, execution errors, and evolving user requirements. Existing multi-turn benchmarks fall short of capturing this complexity, either by treating conversation histories as static context or by limiting evaluation to narrow, read-only (SELECT-ONLY) operations, thereby failing to reflect the challenges encountered in production-grade database assistant. In this work, we introduce BIRD-INTERACT, a benchmark that restores this missing realism through: (1) a ***comprehensive interaction environment*** that couples each database with a hierarchical knowledge base, metadata files, and a function-driven user simulator, enabling models to solicit clarifications, retrieve knowledge, and recover from execution errors without human supervision; (2) two ***evaluation settings*** reflecting real-world interaction settings which contain a pre-defined conversational protocol (c-Interact) and a more open-ended agentic setting (a-Interact) in which the model autonomously decides when to query the user simulator or explore the DB environment; (3) a ***challenging task suite*** that covers the full CRUD spectrum for both business-intelligence and operational use cases, guarded by executable test cases. Each task features ambiguous and follow-up sub-tasks, requiring LLMs to engage in dynamic interaction. The suite is organized into two sets: a full set (**BIRD-INTERACT-FULL**) of 600 tasks which unfold up to **11,796** dynamic interactions for a comprehensive overview of performance and a lite set (**BIRD-INTERACT-LITE**) of 300 tasks, with simplified databases for detailed behavioral analysis of interactions, and fast development of methods. Our empirical results highlight the difficulty of BIRD-INTERACT: the most recent flagship model GPT-5 completes only **8.67%** of tasks in the c-Interact setting and **17.00%** in the a-Interact setting on the full task suite. Further analysis via memory grafting and Interaction Test-time Scaling (ITS), validate the importance of effective interaction for achieving success in complex, dynamic text-to-SQL tasks.

Semivalue–based data valuation uses cooperative‐game theory intuitions to assign each data point a value reflecting its contribution to a downstream task. Still, those values depend on the practitioner’s choice of utility, raising the question: *How robust is semivalue-based data valuation to changes in the utility?* This issue is critical when the utility is set as a trade‐off between several criteria and when practitioners must select among multiple equally valid utilities. We address it by introducing the notion of a dataset’s *spatial signature*: given a semivalue, we embed each data point into a lower-dimensional space where any utility becomes a linear functional, making the data valuation framework amenable to a simpler geometric picture. Building on this, we propose a practical methodology centered on an explicit robustness metric that informs practitioners whether and by how much their data valuation results will shift as the utility changes. We validate this approach across diverse datasets and semivalues, demonstrating strong agreement with rank‐correlation analyses and offering analytical insight into how choosing a semivalue can amplify or diminish robustness.