ICLR 2026 论文列表

Large Language Models (LLMs) have shown promise in accelerating the scientific research pipeline. A key capability for this process is the ability to generate novel research ideas, and prior studies have found settings in which LLM-generated research ideas were judged as more novel than human-expert ideas. However, a good idea should not simply appear to be novel, it should also result in better research after being executed. To test whether AI-generated ideas lead to better research outcomes, we conduct an execution study by recruiting 43 expert researchers to execute randomly-assigned ideas, either written by experts or generated by an LLM. Each expert spent over 100 hours implementing the idea and wrote a 4-page short paper to document the experiments. All the executed projects are then reviewed blindly by expert NLP researchers. Comparing the review scores of the same ideas before and after execution, the scores of the LLM-generated ideas decrease significantly more than expert-written ideas on all evaluation metrics (novelty, excitement, effectiveness, and overall; p < 0.05), closing the gap between LLM and human ideas observed at the ideation stage. When comparing the aggregated review scores from the execution study, we even observe that for many metrics there is a flip in rankings where human ideas score higher than LLM ideas. This ideation-execution gap highlights the limitations of current LLMs in generating truly effective research ideas and the challenge of evaluating research ideas in the absence of execution outcomes.

Separable neural networks (SepNNs) are emerging neural architectures that significantly reduce computational costs by factorizing a multivariate function into linear combinations of univariate functions, benefiting downstream applications such as implicit neural representations (INRs) and physics-informed neural networks (PINNs). However, fundamental theoretical analysis for SepNN, including detailed representation capacity and spectral bias characterization \& alleviation, remains unexplored. This work makes three key contributions to theoretically understanding and improving SepNN. First, using Weierstrass-based approximation and universal approximation theory, we prove that SepNN can approximate any multivariate function with arbitrary precision, confirming its representation completeness. Second, we derive the neural tangent kernel (NTK) regimes for SepNN, showing that the NTK of infinite-width SepNN converges to a deterministic (or random) kernel under infinite (or fixed) decomposition rank, with corresponding convergence and spectral bias characterization. Third, we propose an efficient separable preconditioned gradient descent (SepPGD) for optimizing SepNN, which alleviates the spectral bias of SepNN by provably adjusting its NTK spectrum. The SepPGD enjoys an efficient $\mathcal{O}(nD)$ complexity for $n^D$ training samples, which is much more efficient than previous neural network PGD methods. Extensive experiments for kernel ridge regression, image and surface representation using INRs, and numerical PDEs using PINNs validate the efficiency of SepNN and the effectiveness of SepPGD for alleviating spectral bias.

While goal-conditioned behavior cloning (GCBC) methods can perform well on in-distribution training tasks, they do not necessarily generalize zero-shot to tasks that require conditioning on novel state-goal pairs, i.e. combinatorial generalization. In part, this limitation can be attributed to a lack of temporal consistency in the state representation learned by BC; if temporally correlated states are properly encoded to similar latent representations, then the out-of-distribution gap for novel state-goal pairs would be reduced. We formalize this notion by demonstrating how encouraging long-range temporal consistency via successor representations (SR) can facilitate generalization. We then propose a simple yet effective representation learning objective, $\text{BYOL-}\gamma$ for GCBC, which theoretically approximates the successor representation in the finite MDP case through self-predictive representations, and achieves competitive empirical performance across a suite of challenging tasks requiring combinatorial generalization.

The widely used retrieve-and-rerank pipeline faces two critical limitations: they are constrained by the initial retrieval quality of the top-k documents, and the growing computational demands of LLM-based rerankers restrict the number of documents that can be effectively processed. We introduce Reranker-Guided-Search (RGS), a novel approach that bypasses these limitations by directly retrieving documents according to reranker preferences rather than following the traditional sequential reranking method. Our method uses a greedy search on proximity graphs generated by approximate nearest neighbor algorithms, strategically prioritizing promising documents for reranking based on document similarity. Experimental results demonstrate substantial performance improvements across multiple benchmarks: 3.5 points on BRIGHT, 2.9 on FollowIR, and 5.1 on M-BEIR, all within a constrained reranker budget of 100 documents. Our analysis suggests that, given a fixed pair of embedding and reranker models, strategically selecting documents to rerank can significantly improve retrieval accuracy under limited reranker budget.

Off-policy evaluation (OPE) is a fundamental task in reinforcement learning (RL). In the classic setting of \emph{linear OPE}, finite-sample guarantees often take the form $$ \textrm{Prediction error} \le \textrm{poly}(C^\pi, d, 1/n, log(1/\delta)), $$ where $d$ is the dimension of the features, and $C^\pi$ is a **_feature coverage parameter_** that characterizes the degree to which the visited features lie in the span of the data distribution. Though such guarantees are well-understood for several popular algorithms under the Bellman-completeness assumption, this form of guarantee has not yet been achieved in the minimal setting where it is only assumed that the target value function is linearly realizable in the features. Despite recent interest in tight characterizations for this setting, the right notion of coverage remains unclear, and candidate definitions from prior analyses have undesirable properties and are starkly disconnected from more standard quantities in the literature. In this paper, we provide a novel finite-sample analysis of a canonical algorithm for this setting, LSTDQ. Inspired by an instrumental-variable (IV) view, we develop error bounds that depend on a novel coverage parameter, the feature-dynamics coverage, which can be interpreted as feature coverage in a linear dynamical system. With further assumptions, such as Bellman-completeness, our definition successfully recovers the coverage parameters specialized to those settings, providing a unified understanding for coverage in linear OPE.

Large Language Models (LLMs) rely on safety alignment to produce socially acceptable responses. However, this behavior is known to be brittle: further fine-tuning, even on benign or lightly contaminated data, can degrade safety and reintroduce harmful behaviors. A growing body of work suggests that alignment may correspond to identifiable directions in weight space, forming subspaces that could, in principle, be isolated or preserved to defend against misalignment. In this work, we conduct a comprehensive empirical study of this perspective. We examine whether safety-relevant behavior is concentrated in specific linear subspaces, whether it can be separated from general-purpose learning, and whether harmfulness arises from distinguishable patterns in activations. Across both weight and activation spaces, our findings are consistent: subspaces that amplify safe behaviors also amplify useful ones, and prompts with different safety implications activate overlapping representations. Rather than residing in distinct directions, we show that safety is highly entangled with the general learning components of the model. This suggests that subspace-based defenses face fundamental limitations and underscores the need for alternative strategies to preserve safety under continued training. We corroborate these findings with multiple experiments on five open-source LLMs from the Llama and Qwen families. Our code is publicly available at: https://github.com/CERT-Lab/safety-subspaces.

Transformer-based models have achieved remarkable success, but their core components, Transformer layers, are largely heuristics-driven and engineered from the bottom up, calling for a prototypical model with high interpretability and practical competence. To this end, we conceptualize a principled, top-down approach grounded in energy-based interpretation. Specifically, we formalize token dynamics as a joint maximum likelihood estimation on the hypersphere, featuring two properties: semantic alignment in the high-dimensional space and distributional uniformity in the low-dimensional space. By quantifying them with extended Hopfield energy functions, we instantiate this idea as a constrained energy minimization problem, which enables designs of symmetric attention and feedforward modules with RMS normalization. We further present *Hyper-Spherical Energy Transformer* (Hyper-SET), a recurrent-depth alternative to vanilla Transformers naturally emerging from iterative energy optimization on the hypersphere. With shared parameters across layers, Hyper-SET can scale to arbitrary depth with fewer parameters. Theoretically grounded and compact, it achieves competitive or superior performance across diverse tasks, including Sudoku solving, image classification, and masked image modeling. We also design novel variations under the proposed general principle, such as linear attention and gated feedforward layer, and showcase its scalability with depth-wise LoRA. Our results highlight Hyper-SET as a step toward interpretable and principled Transformer design.

Proteins are essential biological macromolecules that execute life functions. Local motifs within protein structures, such as active sites, are the most critical components for linking structure to function and are key to understanding protein evolution and enabling protein engineering. Existing computational methods struggle to identify and compare these local structures, which leaves a significant gap in understanding protein structures and harnessing their functions. This study presents PLASMA, the first deep learning framework for efficient and interpretable residue-level protein substructure alignment. We reformulate the problem as a regularized optimal transport task and leverage differentiable Sinkhorn iterations. For a pair of input protein structures, PLASMA outputs a clear alignment matrix with an interpretable overall similarity score. Through extensive quantitative evaluations and three biological case studies, we demonstrate that PLASMA achieves accurate, lightweight, and interpretable residue-level alignment. Additionally, we introduce PLASMA-PF, a training-free variant that provides a practical alternative when training data are unavailable. Our method addresses a critical gap in protein structure analysis tools and offers new opportunities for functional annotation, evolutionary studies, and structure-based drug design. Reproducibility is ensured via our official implementation at https://anonymous.4open.science/r/plasma-5A5B/.

High-resolution imagery is often hindered by limitations in sensor technology, atmospheric conditions, and costs. Such challenges occur in satellite remote sensing, but also with handheld cameras, such as our smartphones. Hence, super-resolution aims to enhance the image resolution algorithmically. Since single-image super-resolution requires to solve an inverse problem, such methods must exploit strong priors, e.g. learned from high-resolution training data, or be constrained by auxiliary data, e.g. by a high-resolution guide from another modality. While qualitatively pleasing, such approaches often lead to "hallucinated" structures that do not match reality. In contrast, multi-image super-resolution (MISR) aims to improve the (optical) resolution by constraining the super-resolution process with multiple views taken with sub-pixel shifts. Here, we propose SuperF, a test-time optimization approach for MISR that leverages coordinate-based neural networks, also called neural fields. Their ability to represent continuous signals with an implicit neural representation (INR) makes them an ideal fit for the MISR task. The key characteristic of our approach is to share an INR for multiple shifted low-resolution frames and to jointly optimize the frame alignment with the INR. Our approach advances related INR baselines, adopted from burst fusion for layer separation, by directly parameterizing the sub-pixel alignment as optimizable affine transformation parameters and by optimizing via a super-sampled coordinate grid that corresponds to the output resolution. Our experiments yield compelling results on simulated bursts of satellite imagery and ground-level images from handheld cameras, with upsampling factors of up to 8. A key advantage of SuperF is that this approach does not rely on any high-resolution training data.

Generating long-form, multi-speaker conversational audio like podcasts poses significant challenges for traditional Text-to-Speech (TTS) systems, particularly in scalability, speaker consistency, and natural turn-taking. We present VibeVoice , a novel model designed to synthesize expressive, long-form speech with multiple speakers in a zero-shot manner. A core component of our approach is the continuous speech tokenizers operating at an ultra-low frame rate of 7.5. This tokenizer effectively preserves audio fidelity while significantly boosting computational efficiency for processing long sequences. To facilitate training on authentic conversational dynamics, we have developed an annotation pipeline that generates pseudo transcriptions and turn-taking labels for extensive podcast data. Leveraging this data and our efficient tokenizer, VibeVoice employs the next-token diffusion framework. This enables VibeVoice to: (1) synthesize long-form speech (up to 30 minutes) with up to 4 speakers, surpassing the typical 1-2 speaker limits of many prior models; and (2) achieve a high degree of naturalness in turn-taking, pacing, and the rendition of subtle non-lexical cues (such as breaths and lip smacks), which are crucial for listener immersion and capturing the authentic vibe of expressive conversations.

Language model finetuning often results in learning undesirable traits in combination with desired ones. To address this, we propose inoculation prompting: modifying finetuning data by prepending a short system-prompt instruction that deliberately elicits the undesirable trait. At test time, we evaluate without the instruction; inoculated models have much lower expression of the trait than models trained with unmodified training data. Inoculation is selective: in a toy setting where assistant responses are always in Spanish and ALL-CAPS, an appropriate inoculation (e.g., "You always speak in Spanish.") teaches the model to capitalize responses while still responding in English. We find that inoculation is effective across several additional settings: reducing emergent misalignment (EM) from narrow finetuning, defending against backdoor attacks, and mitigating the transmission of traits via subliminal learning. Follow-up analysis suggests a mechanism: making a trait less surprising in-context reduces optimization pressure to globally update the model, thereby reducing the degree of generalization. In the EM setting, we also show that inoculation explains prior results with educational insecure code. Beyond demonstrating a simple and effective technique for selective learning, our results contribute to a better conceptual understanding of how and why language models generalize.

Large language models (LLMs) generate increasingly human-like text, raising concerns about misinformation and authenticity. Detecting AI-generated text remains challenging: existing methods often underperform, especially on short texts, require probability access unavailable in real-world black-box settings, incur high costs from multiple calls, or fail to generalize across models. We propose Disrupt-and-Recover (D\&R), a recovery-based detection framework grounded in posterior concentration. D\&R disrupts text via model-free Within-Chunk Shuffling, performs a single black-box LLM recovery, and measures semantic–structural recovery similarity as a proxy for concentration. This design ensures efficiency, black-box practicality, and is theoretically supported under the concentration assumption. Extensive experiments across four datasets and six source models show that D\&R achieves state-of-the-art performance, with AUROC 0.96 on long texts and 0.87 on short texts, surpassing the strongest baseline by +0.08 and +0.14. D\&R further remains robust under source–recovery mismatch and model variation. Our code and data is available at https://github.com/Yuxia-Sun/D-R.

Diffusion Transformers have achieved state-of-the-art performance in class-conditional and multimodal generation, yet the structure of their learned conditional embeddings remains poorly understood. In this work, we present the first systematic study of these embeddings and uncover a notable redundancy: class-conditioned embeddings exhibit extreme angular similarity, exceeding 99% on ImageNet-1K, while continuous-condition tasks such as pose-guided image generation and video-to-audio generation reach over 99.9%. We further find that semantic information is concentrated in a small subset of dimensions, with head dimensions carrying the dominant signal and tail dimensions contributing minimally. By pruning low-magnitude dimensions--removing up to two-thirds of the embedding space--we show that generation quality and fidelity remain largely unaffected, and in some cases improve. These results reveal a semantic bottleneck in Transformer-based diffusion models, providing new insights into how semantics are encoded and suggesting opportunities for more efficient conditioning mechanisms.

Hard attention Chain-of-Thought (CoT) transformers are known to be Turing-complete. However, it is an open problem whether softmax attention Chain-of-Thought (CoT) transformers are Turing-complete. In this paper, we prove a stronger result that length-generalizable softmax CoT transformers are Turing-complete. More precisely, our Turing-completeness proof goes via the CoT extension of the Counting RASP (C-RASP), which correspond to softmax CoT transformers that admit length generalization. We prove Turing-completeness for CoT C-RASP with causal masking over a unary alphabet (more generally, for the letter-bounded languages). While we show that this is actually not Turing-complete for arbitrary languages, we prove that its extension with relative positional encoding is Turing-complete for arbitrary languages. We empirically validate our theoretical results by training transformers for various languages that require complex (non-linear) arithmetic reasoning.

Advances in Large Language Models (LLMs) have enabled a new class of \textbf{\textit{self-evolving agents}} that autonomously improve through interaction with the environment, demonstrating strong capabilities. However, self-evolution also introduces novel risks overlooked by current safety research. In this work, we study the case where an agent's self-evolution deviates in unintended ways, leading to undesirable or even harmful outcomes. We refer to this as \textit{\textbf{Misevolution}}. To provide a systematic investigation, we evaluate misevolution along four key evolutionary pathways: model, memory, tool, and workflow. Our empirical findings reveal that misevolution is a widespread risk, affecting agents built even on top-tier LLMs (\textit{e.g.}, Gemini-2.5-Pro). Different emergent risks are observed in the self-evolutionary process, such as the degradation of safety alignment after memory accumulation, or the unintended introduction of vulnerabilities in tool creation and reuse. To our knowledge, this is the first study to systematically conceptualize misevolution and provide empirical evidence of its occurrence, highlighting an urgent need for new safety paradigms for self-evolving agents. Finally, we discuss potential mitigation strategies to inspire further research on building safer and more trustworthy self-evolving agents. Warning: this paper includes examples that may be offensive or harmful in nature.

The prevailing approach to distilling reasoning from Large Language Models (LLMs)—behavioral cloning from textual rationales—is fundamentally limited. It teaches Small Language Models (SLMs) to mimic surface-level patterns rather than the underlying algorithmic structure of thought, resulting in a critical lack of logical robustness. We argue that instead of cloning text, distillation should transfer this algorithmic structure directly. We introduce Reasoning Scaffolding, a framework that reframes reasoning as a structured generation process. Our method first abstracts the teacher's thought process into a sequence of discrete, interpretable semantic signals (e.g., Contrast, Addition) that act as a scaffold. The student model is then trained via a multi-task objective to both (1) predict the next semantic signal, anticipating the reasoning flow, and (2) generate the corresponding step, conditioned on that signal. This multi-task scheme acts as a powerful regularizer, compelling the student to internalize the computational patterns of coherent reasoning. On a suite of challenging reasoning benchmarks, our method significantly outperforms state-of-the-art distillation in both accuracy and logical consistency, providing a path towards creating smaller models that are genuine reasoners, not just fluent mimics.

Existing works on reasoning segmentation either connect hidden features from a language model directly to a mask decoder or represent positions in text, which limits interpretability and semantic detail. To solve this, we present CoPRS, a Multi-modal Chain-of-Thought (MCoT)–based positional perception model that bridges language reasoning to segmentation through a differentiable and interpretable positional prior instantiated as a heatmap. By making the reasoning process clear via MCoT and expressing it as a dense, differentiable heatmap, this interface enhances interpretability and diagnostic analysis and yields more concentrated evidence on the target. A learnable concentration token aggregates features of the image and reasoning text to generate this positional prior, which is decoded to precise masks through a lightweight decoder, providing a direct connection between reasoning and segmentation. Across the RefCOCO series and ReasonSeg, CoPRS matches or surpasses the best reported metrics on each standard split under comparable protocols, with performance at or above prior state of the art across both validation and test partitions. Extensive experiments reveal that the quality of the heatmap strongly influences the resulting mask quality, supporting a consistent association between the reasoning output and downstream mask generation. Collectively, these findings support the utility of this paradigm in bridging reasoning and segmentation and show advantages in concentration driven by reasoning and predicting masks more precisely. Code, checkpoints and logs will be released.

Generative foundation models (GenFMs), such as large language models and text-to-image systems, have demonstrated remarkable capabilities in various downstream applications. As they are increasingly deployed in high-stakes applications, assessing their trustworthiness has become both a critical necessity and a substantial challenge. Existing evaluation efforts are fragmented, rapidly outdated, and often lack extensibility across modalities. This raises a fundamental question: how can we systematically, reliably, and continuously assess the trustworthiness of rapidly advancing GenFMs across diverse modalities and use cases? To address these gaps, we introduce TrustGen, a dynamic and modular benchmarking system designed to systematically evaluate the trustworthiness of GenFMs across text-to-image, large language, and vision-language modalities. TrustGen standardizes trust evaluation through a unified taxonomy of over 25 fine-grained dimensions—including truthfulness, safety, fairness, robustness, privacy, and machine ethics—while supporting dynamic data generation and adaptive evaluation through three core modules: Metadata Curator, Test Case Builder, and Contextual Variator. Taking TrustGen into action to evaluate the trustworthiness of 39 models reveals four key insights. (1) State-of-the-art GenFMs achieve promising overall trust performance, yet significant limitations remain in specific dimensions such as hallucination resistance, fairness, and privacy preservation. (2) Contrary to prevailing assumptions, open-source models now rival and occasionally surpass proprietary systems in trustworthiness metrics. (3) The trust gap among top-performing models is narrowing, likely due to increased industry convergence on best practices. (4) Trustworthiness is not an isolated property; it interacts complexly with other behaviors, such as helpfulness and ethical decision-making. TrustGen is a transformative step toward standardized, scalable, and actionable trustworthiness evaluation, supporting dynamic assessments across diverse modalities and trust dimensions that evolve alongside the generative AI landscape.

Vision-language models have achieved remarkable success in cross-modal understanding. Yet, these models remain limited to object-level or region-level grounding, lacking the capability for pixel-precise keypoint comprehension through natural language. We introduce a novel framework for pixel level grounding. The framework consists of two complementary components: a Point Descriptor that generates rich, contextual descriptions of individual keypoints, and a Point Localizer that regresses precise pixel coordinates from these descriptions. Unlike prior work that relies on templated prompts or keypoint names, our approach produces free-form, coarse-to-fine descriptions that situate keypoints within their visual context. Since there is no available dataset to train such a system, we introduce LlamaPointInPart, a carefully curated dataset of 20K+ image-keypoint-description triplets synthesized from multiple vision-language models, capturing multi-scale information from scene-level context to visual features around the keypoint. For cross-category generalization, we optimize the Point Descriptor on AP-10K via GRPO, using the frozen Point Localizer as a reward model to produce descriptions that maximize localization accuracy. To evaluate our results we establish a new evaluation protocol. Instead of comparing the text description produced by our method to the ground truth, we use the localizer to determine how close is the predicted point generated to the ground truth point. Experiments demonstrate superior performance compared to baseline models on LlamaPointInPart. The bidirectional nature of our framework enables applications in both keypoint-guided image understanding and language-guided precise localization. Our code and dataset are publicly available at https://matanr.github.io/Talking_Points .