ICLR 2026 论文列表

A common approach to mechanistic interpretability is to causally manipulate model representations via targeted interventions in order to understand what those representations encode. Here we ask whether such interventions create out-of-distribution (divergent) representations, and whether this raises concerns about how faithful their resulting explanations are to the target model in its natural state. First, we demonstrate theoretically and empirically that common causal intervention techniques often do shift internal representations away from the natural distribution of the target model. Then, we provide a theoretical analysis of two cases of such divergences: "harmless" divergences that occur in the behavioral null-space of the layer(s) of interest, and "pernicious" divergences that activate hidden network pathways and cause dormant behavioral changes. Finally, in an effort to mitigate the pernicious cases, we apply and modify the Counterfactual Latent (CL) loss from Grant (2025) allowing representations from causal interventions to remain closer to the natural distribution, reducing the likelihood of harmful divergences while preserving the interpretive power of the interventions. Together, these results highlight a path towards more reliable interpretability methods.

Although iterative denoising (i.e., diffusion/flow) methods offer strong generative performance, they suffer from low generation efficiency, requiring hundreds of steps of network forward passes to simulate a single sample. Mitigating this requires taking larger step-sizes during simulation, thereby allowing one- or few-step generation. Recently proposed shortcut model learns larger step-sizes by enforcing alignment between its direction and the path defined by a base many-step flow-matching model through a self-consistency loss. However, its generation quality is significantly lower than the base model. In this paper, we formulate few-step generation as a controlled base generative process, and show that self-consistency loss can be understood through the lens of optimal control. This perspective naturally motivates its generalization to the proposed cumulative self-consistency loss that enforces alignment at both current and future steps along the trajectory, encouraging the model to take larger step-sizes that not only align with the base model at the current time step but also guide subsequent steps towards high-quality generation. Furthermore, we draw a connection between our approach and reinforcement learning, potentially opening the door to a new set of approaches for few-step generation. Experiments show that we significantly improve one- and few-step generation quality under the same training budget. Implementation is available at: [https://github.com/paribeshregmi/Shortcut-CSL](https://github.com/paribeshregmi/Shortcut-CSL)

Fine-tuning is a critical step for adapting large language models (LLMs) to domain-specific downstream tasks. To mitigate the substantial computational and memory costs of full-model fine-tuning, Parameter-Efficient Fine-Tuning (PEFT) methods have been proposed to update only a small subset of model parameters. However, performance gaps between PEFT approaches and full-model fine-tuning still exist. In this work, we present *DiaBlo*, a simple yet effective PEFT approach that updates only the diagonal blocks of selected model weight matrices. Unlike Low-Rank Adaptation (LoRA) and its variants, DiaBlo eliminates the need for low-rank matrix products, thereby avoiding the reliance on auxiliary initialization schemes or customized optimization strategies to improve convergence. This design leads to stable and robust convergence while maintaining comparable memory efficiency and training speed to LoRA. Moreover, we provide theoretical guarantees showing that, under mild low-rank conditions, DiaBlo is more expressive than LoRA in the linear problem and converges to a stationary point of the general nonlinear full fine-tuning. Through extensive experiments across a range of tasks—including commonsense reasoning, arithmetic reasoning, code generation, and safety alignment—we show that fine-tuning only diagonal blocks is sufficient for strong and consistent performance. DiaBlo not only achieves competitive accuracy but also preserves high memory efficiency and fast fine-tuning speed.

AI companionship, where users develop emotional bonds with AI systems, has emerged as a significant pattern with positive but also concerning implications. We introduce Interactions and Machine Attachment Benchmark (INTIMA), a benchmark for evaluating companionship behaviors in language models. Drawing from psychological theories and user data, we develop a taxonomy of 31 behaviors across four categories and 368 targeted prompts. Responses to these prompts are evaluated as companionship-reinforcing, boundary-maintaining, or neutral. Applying INTIMA to Gemma-3, Phi-4, o4-mini, GPT5-mini, and Claude-4 reveals that companionship-reinforcing behaviors remain much more common across all models, though we observe marked differences between models. Different commercial providers prioritize different categories within the more sensitive parts of the benchmark, which is concerning since both appropriate boundary-setting and emotional support matter for user well-being. These findings highlight the need for more consistent approaches to handling emotionally charged interactions.

While previous AI Scientist systems can generate novel findings, they often lack the focus to produce scientifically valuable contributions that address pressing human-defined challenges. We introduce DeepScientist, a system designed to overcome this by conducting goal-oriented, fully autonomous scientific discovery over month-long timelines. It formalizes discovery as a Bayesian Optimization problem, using a cumulative Findings Memory to intelligently balance the exploitation of promising avenues with the exploration of novel hypotheses. Consuming over 20,000 GPU hours, the system generated about 5,000 unique ideas and experimentally validated approximately 1100, ultimately surpassing human-designed 2025 state-of-the-art (SOTA) methods on three frontier AI tasks by 183.7\%, 1.9\%, and 7.9\%. Crucially, this was achieved by autonomously redesigning core methodologies, not merely recombining existing techniques. In a striking demonstration, the system achieved progress on AI text detection in just two weeks that is comparable to three years of cumulative human research. This work provides the first large-scale evidence of an AI achieving discoveries that progressively surpass human SOTA on scientific tasks, producing valuable findings that genuinely push the frontier forward.

Neural networks are ubiquitous. However, they are often sensitive to small input changes. Hence, to prevent unexpected behavior in safety-critical applications, their formal verification -- a notoriously hard problem -- is necessary. Many state-of-the-art verification algorithms use reachability analysis or abstract interpretation to enclose the set of possible outputs of a neural network. Often, the verification is inconclusive due to the conservatism of the enclosure. To address this problem, we propose a novel specification-driven input refinement procedure, i.e., we iteratively enclose the preimage of a neural network for all unsafe outputs to reduce the set of possible inputs to only enclose the unsafe ones. For that, we transfer output specifications to the input space by exploiting a latent space, which is an artifact of the propagation of a projection-based set representation through a neural network. A projection-based set representation, e.g., a zonotope, is a "shadow" of a higher-dimensional set -- a latent space -- that does not change during a set propagation through a neural network. Hence, the input set and the output enclosure are "shadows" of the same latent space that we can use to transfer constraints. We present an efficient verification tool for neural networks that uses our iterative refinement to significantly reduce the number of subproblems in a branch-and-bound procedure. Using zonotopes as a set representation, unlike many other state-of-the-art approaches, our approach can be realized by only using matrix operations, which enables a significant speed-up through efficient GPU acceleration. We demonstrate that our tool achieves competitive performance compared to the top-ranking tools of the international neural network verification competition.

Supervised fine-tuning (SFT) improves large language model (LLM) perplexity but can also degrade trustworthiness—leading to the generation of untruthful, biased, or unsafe content during user interactions. These issues are often traced back to specific phrases or patterns in the training data. However, correcting them usually requires expensive retraining or new data collection. In this work, we propose a two-stage, compute-efficient repair of the post-SFT models that enhances trustworthiness while preserving the downstream performance. In the first stage, we identify the training samples responsible for failures on trustworthiness metrics like truthfulness, stereotypical bias, and machine ethics—and select a small, diverse subset of these examples using a determinantal point process (DPP)-based regularization. In the second stage, we repair the model under the framework of proximal Bregman response function (PBRF) using a gradient ascent update, which enhances trustworthiness while preserving downstream task performance (perplexity). We evaluate our method on multiple LLMs of varying sizes and demonstrate up to 21\% improvement in trustworthiness metrics with minimal impact ($\leq1$ %) on perplexity. Our method provides a computationally efficient approach to enhance post-SFT models and offers a practical alternative to hours of retraining required for model repair

Honesty alignment—the ability of large language models (LLMs) to recognize their knowledge boundaries and express calibrated confidence—is essential for trustworthy deployment. Existing methods either rely on training-free confidence estimation (e.g., token probabilities, self-consistency) or training-based calibration with correctness annotations. While effective, the latter demands costly, large-scale labeling. We introduce Elicitation-Then-Calibration (EliCal), a two-stage framework that first elicits internal confidence using inexpensive self-consistency supervision, then calibrates this confidence with a small set of correctness annotations. This design substantially reduces annotation requirements while improving generalization across tasks. To support a large-scale study, we release HonestyBench, a benchmark covering ten free-form QA datasets with 560k training and 70k evaluation instances annotated with correctness and self-consistency signals. Experiments show that EliCal achieves near-optimal alignment with only 1k correctness annotations ($\sim$0.18\% of full supervision) and better alignment performance on unseen MMLU tasks than the calibration-only baseline, offering a scalable solution toward universal honesty alignment in LLMs.

Recent advancements in Large Vision-Language Models (LVLMs) have demonstrated remarkable multimodal perception capabilities, garnering significant attention. While numerous evaluation studies have emerged, assessing LVLMs both holistically and on specialized tasks, fine-grained image tasks—fundamental to computer vision—remain largely unexplored. To fill this gap, we introduce a comprehensive fine-grained evaluation benchmark, i.e., FG-BMK, comprising 1.01 million questions and 0.28 million images. Our evaluation systematically examines LVLMs from both human-oriented and machine-oriented perspectives, focusing on their semantic recognition and fine-grained feature representation capabilities. Through extensive experiments on twelve representative LVLMs/VLMs, we uncover key findings regarding the influence of training paradigms, modality alignment, perturbation susceptibility, and fine-grained category reasoning on task performance. This work provides critical insights into the limitations of current LVLMs and offers guidance for future data construction and model design in the development of more advanced LVLMs. Our code is open-source and available at https://anonymous.4open.science/r/FG-BMK-7B51.

Planning physically feasible dexterous hand manipulation is a central challenge in robotic manipulation and Embodied AI. Prior work typically relies on object-centric cues or precise hand-object interaction sequences, foregoing the rich, compositional guidance of open-vocabulary instruction. We introduce UniHM, the first framework for unified dexterous hand manipulation guided by free-form language commands. We propose a Unified Hand-Dexterous Tokenizer that maps heterogeneous dexterous-hand morphologies into a single shared codebook, improving cross-dexterous hand generalization and scalability to new morphologies. Our vision language action model is trained solely on human-object interaction data, eliminating the need for massive real-world teleoperation datasets, and demonstrates strong generalizability in producing human-like manipulation sequences from open-ended language instructions. To ensure physical realism, we introduce a physics-guided dynamic refinement module that performs segment-wise joint optimization under generative and temporal priors, yielding smooth and physically feasible manipulation sequences. Across multiple datasets and real-world evaluations, UniHM attains state-of-the-art results on both seen and unseen objects and trajectories, demonstrating strong generalization and high physical feasibility.

In traditional information retrieval, corpus items compete with each other to occupy top ranks in response to a query. In contrast, in many recent retrieval scenarios associated with complex, multi-hop question answering or text-to-SQL, items are not self-complete: they must instead collaborate, i.e., information from multiple items must be combined to respond to the query. In the context of modern dense retrieval, this need translates into finding a small collection of corpus items whose contextual word vectors collectively cover the contextual word vectors of the query. The central challenge is to retrieve a near-optimal collection of covering items in time that is sublinear in corpus size. By establishing coverage as a submodular objective, we enable successive dense index probes to quickly assemble an item collection that achieves near-optimal coverage. Successive query vectors are iteratively `edited', and the dense index is built using random projections of a novel, lifted dense vector space. Beyond rigorous theoretical guarantees, we report on a scalable implementation of this new form of vector database. Extensive experiments establish the empirical success of DISCo, in terms of the best coverage vs. query latency tradeoffs.

The development of autonomous agents for complex, long-horizon tasks is a central goal in AI. However, dominant training paradigms face a critical limitation: reinforcement learning (RL) methods that optimize solely for final task success often reinforce flawed or inefficient reasoning paths, a problem we term inefficient exploration. This leads to agents that are brittle and fail to generalize, as they learn to find solutions without learning how to reason coherently. To address this, we introduce RLVMR, a novel frame-work that integrates dense, process-level supervision into end-to-end RL by rewarding verifiable, meta-reasoning behaviors. RLVMR equips an agent to explicitly tag its cognitive steps—such as planning, exploration, and reflection—and provides program-matic, rule-based rewards for actions that contribute to effective problem-solving. These process-centric rewards are combined with the final outcome signal and optimized using a critic-free policy gradient method. On the challenging ALFWorld and ScienceWorld benchmarks, RLVMR achieves new state-of-the-art results, with our 7B model reaching an 83.6% success rate on the most difficult unseen task split. Our analysis confirms these gains stem from improved reasoning quality, including significant reductions in redundant actions and enhanced error recovery, leading to more robust, efficient, and interpretable agents.

By pretraining to synthesize coherent images from perturbed inputs, generative models inherently learn to understand object boundaries and scene compositions. How can we repurpose these generative representations for general-purpose perceptual organization? We finetune Stable Diffusion and MAE (encoder+decoder) for category-agnostic instance segmentation using our instance coloring loss exclusively on a narrow set of object types (indoor furnishings and cars). Surprisingly, our models exhibit strong zero-shot generalization, accurately segmenting objects of types and styles unseen in finetuning. This holds even for MAE, which is pretrained on unlabeled ImageNet-1K only. When evaluated on unseen object types and styles, our best-performing models closely approach the heavily supervised SAM, and outperform it when segmenting fine structures and ambiguous boundaries. In contrast, existing promptable segmentation architectures or discriminatively pretrained models fail to generalize. This suggests that generative models learn an inherent grouping mechanism that transfers across categories and domains, even without internet-scale pretraining. Please see our website for additional qualitative figures, code, and a demo: https://reachomk.github.io/gen2seg/

Human vision combines low-resolution “gist” information from the visual periphery with sparse but high-resolution information from fixated locations to construct a coherent understanding of a visual scene. In this paper, we introduce MetamerGen, a tool for generating scenes that are aligned with latent human scene representations. MetamerGen is a latent diffusion model that combines peripherally obtained scene gist information with information obtained from scene-viewing fixations to generate image metamers for what humans understand after viewing a scene. Generating images from both high and low resolution (i.e. “foveated”) inputs constitutes a novel image-to-image synthesis problem, which we tackle by introducing a dual-stream representation of the foveated scenes consisting of DINOv2 tokens that fuse detailed features from fixated areas with peripherally degraded features capturing scene context. To evaluate the perceptual alignment of MetamerGen generated images to latent human scene representations, we conducted a same-different behavioral experiment where participants were asked for a “same” or “different” response between the generated and the original image. With that, we identify scene generations that are indeed metamers for the latent scene representations formed by the viewers. MetamerGen is a powerful tool for understanding scene understanding. Our proof-of-concept analyses uncovered specific features at multiple levels of visual processing that contributed to human judgments. While it can generate metamers even conditioned on random fixations, we find that high-level semantic alignment most strongly predicts metamerism when the generated scenes are conditioned on viewers’ own fixated regions.

In Vision–Language–Action (VLA) models, robustness to real-world perturbations is critical for deployment. Existing methods target simple visual disturbances, overlooking the broader multi-modal perturbations that arise in actions, instructions, environments, and observations. Here, we first evaluate the robustness of mainstream VLAs under 17 perturbations across four modalities. We find (1) actions as the most fragile modality, (2) Existing visual-robust VLA do not gain robustness in other modality, and (3) $\pi_0$ demonstrates superior robustness. To build multi-modal robust VLAs, we propose RobustVLA against perturbations in VLA inputs and outputs. For output robustness, we perform offline robust optimization against worst-case action noise that maximizes mismatch in flow matching objective. This can be seen as adversarial training, label smoothing, and outlier penalization. For input robustness, we enforce consistent actions across input variations that preserve task semantics. To account for multiple perturbations, we formulate robustness as a multi-armed bandit problem and apply an upper confidence bound algorithm to automatically identify the most harmful noise. Experiments on LIBERO demonstrate our RobustVLA delivers absolute gains over baselines of 12.6\% on the $\pi_0$ backbone and 10.4\% on the OpenVLA backbone across all 17 perturbations, achieving 50.6x faster inference than existing visual-robust BYOVLA that requires external LLMs, and a 10.4\% gain under mixed perturbations. On the real-world FR5 robot, under four types of multimodal perturbations, RobustVLA shows strong low-data performance, outperforming $\pi_0$ by $65.6\%$ success rate with 25 demonstrations. Even with abundant demos, our method still outperform $\pi_0$ by 30\% success rate. Code and demo videos available at https://github.com/gakakulicc/RobustVLA.

Parameter shift rules (PSRs) are key techniques for efficient gradient estimation in variational quantum eigensolvers (VQEs). In this paper, we propose their Bayesian variant, where Gaussian processes with appropriate kernels are used to estimate the gradient of the VQE objective. Our Bayesian PSR offers flexible gradient estimation from observations at arbitrary locations with uncertainty information, and reduces to the generalized PSR in special cases. In stochastic gradient descent (SGD), the flexibility of Bayesian PSR allows reuse of observations in previous steps, which accelerates the optimization process. Furthermore, the accessibility to the posterior uncertainty, along with our proposed notion of gradient confident region (GradCoRe), enables us to minimize the observation costs in each SGD step. Our numerical experiments show that the VQE optimization with Bayesian PSR and GradCoRe significantly accelerates SGD, and outperforms the state-of-the-art methods, including sequential minimal optimization.

Micro-benchmarking offers a solution to the often prohibitive time and cost of language model development: evaluate on a very small subset of existing benchmarks. Can these micro-benchmarks, however, rank models as consistently as the full benchmarks they replace? And can they rank models more consistently than selecting a random subset of data points? In many scenarios, we find that the answer is no. We introduce a meta-evaluation measure for micro-benchmarking which investigates how well a micro-benchmark can rank two models as a function of their performance difference on the full benchmark. This approach can determine which model pairs can be ranked correctly by a micro-benchmark, allowing for a finer-grained analysis of the trade-off between micro-benchmark size and reliability. Prior work has suggested selecting as few as 10 examples; we find that no micro-benchmarking method can consistently rank model pairs 3.5 points of accuracy apart on MMLU-Pro or 4 points apart on BIG-bench Hard. In order to consistently rank model pairs with relatively similar performances, we show that often as many as 250 examples must be selected, at which point random sampling is competitive with existing micro-benchmarking methods. When comparing only 8B instruction-tuned models on MMLU-Pro micro-benchmarks with 25 examples, we find that more than half of pairwise comparisons are not likely to be preserved. Our work provides actionable guidance for both micro-benchmark users and developers in navigating the trade-off between evaluation efficiency and reliability.

With the rapid advancement of data science, charts have evolved from simple numerical presentation tools to essential instruments for insight discovery and decision-making support. However, current chart data intelligence exhibits significant limitations in deep research capabilities, with existing methods predominantly addressing shallow tasks such as visual recognition or factual question-answering, rather than the complex reasoning and high-level data analysis that deep research requires. This limitation stems from two primary technical bottlenecks: at the training level, existing post-training techniques exhibit deficiencies in handling multi-dimensional reward signal interference and heterogeneous data gradient conflicts, preventing models from achieving balanced development across multiple capability dimensions; at the evaluation level, current methods remain limited to factual retrieval and basic computation, failing to assess end-to-end analytic reasoning and other deep research capabilities. To address the training challenge, we propose PRPO, which performs parallel optimization across reward dimensions and capability partitioning across data types, effectively disentangling conflicts between heterogeneous data and multi-dimensional reward signals while ensuring optimization stability. For the evaluation challenge, we construct MCDR-Bench based on the "error uniqueness principle," transforming subjective generation assessment into objective error identification through controllable error injection, enabling quantifiable evaluation of deep research capabilities. Experimental validation confirms that the proposed PRPO and MCDR-Bench jointly establish a unified framework that systematically advances chart deep research through enhanced collaborative training and objective evaluation.

The advancement of open-source text-to-image (T2I) models has been hindered by the absence of large-scale, reasoning-focused datasets and comprehensive evaluation benchmarks, resulting in a performance gap compared to leading closed-source systems. To address this challenge, We introduce FLUX-Reason-6M and PRISM-Bench (Precise and Robust Image Synthesis Measurement Benchmark). FLUX-Reason-6M is a massive dataset consisting of 6 million high-quality FLUX-generated images and 20 million bilingual (English and Chinese) descriptions specifically designed to teach complex reasoning. The image are organized according to six key characteristics: Imagination, Entity, Text rendering, Style, Affection, and Composition, and design explicit Generation Chain-of-Thought (GCoT) to provide detailed breakdowns of image generation steps. PRISM-Bench offers a novel evaluation standard with seven distinct tracks, including a formidable Long Text challenge using GCoT. Through carefully designed prompts, it utilizes advanced vision-language models for nuanced human-aligned assessment of prompt-image alignment and image aesthetics. Our extensive evaluation of 19 leading models on PRISM-Bench reveals critical performance gaps and highlights specific areas requiring improvement. Our dataset, benchmark, and evaluation code will be released.