ICLR 2026 论文列表

In this paper, we address post-training quantization (PTQ) for large language models (LLMs) from an overlooked perspective: given a pre-trained high-precision LLM, the predominant sequential quantization framework treats different layers equally, but this may be not optimal in challenging bit-width settings. We empirically study the quantization impact of different layers on model accuracy, and observe that: (1) shallow/deep layers are usually more sensitive to quantization than intermediate layers; (2) among shallow/deep layers, the most sensitive one is the first/last layer, which exhibits significantly larger quantization error than others. These empirical observations imply that the quantization design for different layers of LLMs is required on multiple levels instead of a single level shared to all layers. Motivated by this, we propose a new PTQ framework termed **Sliding**-lay**er** **Quant**ization (SliderQuant) that relies on a simple adaptive sliding quantization concept facilitated by few learnable parameters. The base component of SliderQuant is called inter-layer sliding quantization, which incorporates three types of novel sliding window designs tailored for addressing the varying quantization sensitivity of shallow, intermediate and deep layers. The other component is called intra-layer sliding quantization that leverages an incremental strategy to quantize each window. As a result, SliderQuant has a strong ability to reduce quantization errors across layers. Extensive experiments on basic language generation, zero-shot commonsense reasoning and challenging math and code tasks with various LLMs, including Llama/Llama2/Llama3/Qwen2.5 model families, DeepSeek-R1 distilled models and large MoE models, show that our method outperforms existing PTQ methods (including the latest PTQ methods using rotation transformations) for both weight-only quantization and weight-activation quantization under diverse bit width settings. Code will be made publicly available.

Large language models (LLMs) benefit from test-time scaling but are often hampered by high inference latency. Speculative decoding is a natural way to accelerate the scaling process; however, scaling along both the parallel and sequential dimensions poses significant challenges, including substantial memory-bound execution and synchronization overhead. We introduce *ATTS* (Asynchronous Test-Time Scaling), a statistically guaranteed adaptive scaling framework that follows the hypothesis testing process to address these challenges. By revisiting arithmetic intensity, *ATTS* identifies synchronization as the primary bottleneck. It enables asynchronous inference through online calibration and proposes an ordinal classification algorithm that supports a three-stage rejection sampling pipeline, scaling along both the sequential and parallel axes. Across experiments on the MATH, AMC23, AIME24, and AIME25 datasets and across multiple draft–target model families, we show that *ATTS* delivers up to *56.7x* speedup in test-time scaling and a *4.14x* throughput improvement, while maintaining accurate control of the rejection rate, reducing latency and memory overhead, and incurring no accuracy loss. By scaling both in parallel and sequential dimensions, we enable the 1.5B/70B draft/target model combination to achieve the performance of the state-of-the-art reasoning model o3-mini (high) on the AIME dataset. We submit the anonymous repository: anonymous.4open.science/r/Asynchronous-Test-Time-Scaling-5940.

Large Language Models (LLMs) exhibit impressive capabilities but often hallucinate, confidently providing incorrect answers instead of admitting ignorance. Prior work has shown that models encode linear representations of their own knowledge and that activation steering can reduce hallucinations. These approaches, however, require real-time monitoring and intervention during inference. We introduce Contrastive Activation Steering for Amortized Learning (CASAL), an efficient algorithm that connects interpretability with amortized optimization. CASAL directly bakes the benefits of activation steering into model's weights. Once trained, LLMs answer questions they know while abstaining from answering those they do not. CASAL's light-weight design requires training only a submodule of a single transformer layer and yet reduces hallucination by $\sim30\%$-$40 \%$ across multiple short-form QA benchmarks. CASAL is $\sim$30x more compute-efficient and $\sim$20x more data-efficient than strong LoRA-based baselines such as SFT and DPO, boosting its practical applicability in data scarce domains. Importantly, CASAL also generalizes effectively to out-of-distribution (OOD) domains. We showcase CASAL's flexibility in mitigating hallucinations in both text-only and vision-language models. To our knowledge, CASAL is the first steering-based training method that has been shown to be effective for both dense and Mixture-of-Experts (MoE) models. CASAL represents a promising step forward for applying interpretability-inspired method for practical deployment in production systems.

Inductive link prediction with knowledge hypergraphs is the task of predicting missing hyperedges involving completely *novel entities* (i.e., nodes unseen during training). Existing methods for inductive link prediction with knowledge hypergraphs assume a fixed relational vocabulary and, as a result, cannot generalize to knowledge hypergraphs with *novel relation types* (i.e., relations unseen during training). Inspired by knowledge graph foundation models, we propose HYPER as a foundation model for link prediction, which can generalize to *any knowledge hypergraph*, including novel entities and novel relations. Importantly, HYPER can learn and transfer across different relation types of *varying arities*, by encoding the entities of each hyperedge along with their respective positions in the hyperedge. To evaluate HYPER, we construct 16 new inductive datasets from existing knowledge hypergraphs, covering a diverse range of relation types of varying arities. Empirically, HYPER consistently outperforms all existing methods in both node-only and node-and-relation inductive settings, showing strong generalization to unseen, higher-arity relational structures.

Recent advancements in diffusion and flow-matching models have demonstrated remarkable capabilities in high-fidelity image synthesis. A prominent line of research involves reward-guided guidance, which steers the generation process during inference to align with specific objectives. However, leveraging this reward-guided approach to the task of image editing, which requires preserving the semantic content of the source image while enhancing a target reward, is largely unexplored. In this work, we introduce a novel framework for training-free, reward-guided image editing. We formulate the editing process as a trajectory optimal control problem where the reverse process of a diffusion model is treated as a controllable trajectory originating from the source image, and the adjoint states are iteratively updated to steer the editing process. Through extensive experiments across distinct editing tasks, we demonstrate that our approach significantly outperforms existing inversion-based training-free guidance baselines, achieving a superior balance between reward maximization and fidelity to the source image without reward hacking.

Current speech-language models (SLMs) typically use a cascade of speech encoder and large language model, treating speech understanding as a single black box. They analyze the content of speech well but reason weakly about other aspects, especially under sparse supervision. Thus, we argue for explicit reasoning over speech states and actions with modular and transparent decisions. Inspired by cognitive science we adopt a modular perspective and a world model view in which the system learns forward dynamics over latent states. We factorize speech understanding into four modules that communicate through a causal graph, establishing a cognitive state search space. Guided by posterior traces from this space, an instruction-tuned language model produces a concise causal analysis and a user-facing response, enabling counterfactual interventions and interpretability under partial supervision. We present the first graph based modular speech model for explicit reasoning and we will open source the model and data to promote the development of advanced speech understanding.

Among on-policy reinforcement learning algorithms, Proximal Policy Optimization (PPO) demonstrates is widely favored for its simplicity, numerical stability, and strong empirical performance. Standard PPO relies on surrogate objectives defined via importance ratios, which require evaluating policy likelihood that is typically straightforward when the policy is modeled as a Gaussian distribution. However, extending PPO to more expressive, high-capacity policy models such as continuous normalizing flows (CNFs), also known as flow-matching models, is challenging because likelihood evaluation along the full flow trajectory is computationally expensive and often numerically unstable. To resolve this issue, we propose PolicyFlow, a novel on-policy CNF-based reinforcement learning algorithm that integrates expressive CNF policies with PPO-style objectives without requiring likelihood evaluation along the full flow path. PolicyFlow approximates importance ratios using velocity field variations along a simple interpolation path, reducing computational overhead without compromising training stability. To further prevent mode collapse and further encourage diverse behaviors, we propose the Brownian Regularizer, an implicit policy entropy regularizer inspired by Brownian motion, which is conceptually elegant and computationally lightweight. Experiments on diverse tasks across vairous environments including MultiGoal, PointMaze, IsaacLab and MuJoCo Playground show that PolicyFlow achieves competitive or superior performance compared to PPO using Gaussian policies and flow-based baselines including FPO and DPPO. Notably, results on MultiGoal highlight PolicyFlow’s ability to capture richer multimodal action distributions.

Classical influence functions face significant challenges when applied to deep neural networks, primarily due to non-invertible Hessians and high-dimensional parameter spaces. We propose the local Bayesian influence function (BIF), an extension of classical influence functions that replaces Hessian inversion with loss landscape statistics that can be estimated via stochastic-gradient MCMC sampling. This Hessian-free approach captures higher-order interactions among parameters and scales efficiently to neural networks with billions of parameters. We demonstrate state-of-the-art results on predicting retraining experiments.

The performance of Large Language Models (LLMs) hinges on carefully engineered prompts. However, prevailing prompt optimization methods, ranging from heuristic edits and reinforcement learning to evolutionary search, primarily target point-wise accuracy. They seldom enforce paraphrase invariance or searching stability, and therefore cannot remedy this brittleness in practice. Automated prompt search remains brittle: small, semantically preserving paraphrases often cause large performance swings. We identify this brittleness as the **textual sharpness** of the **prompt landscape**. In this work, we provide the first formal treatment of textual sharpness in the discrete, semantic space of prompts, together with an operational robustness criterion over a semantic neighborhood; the design is black-box or API-only, requiring no gradients to update the model's parameters. Then we introduce **TARE** (Textual Sharpness-Aware Evolving), a derivative-free framework that alternates between an inner, sampling-based adversarial search that stresses a prompt with hard paraphrases and an outer, robust selection that prefers candidates whose neighborhoods remain strong. We further propose **ATARE**, which learns anisotropic weights to shape the semantic neighborhood and adapts its radius over time to balance exploration and fidelity. Diverse tasks evaluate our methods, whose design for minimizing textual sharpness gap leads to prompts that preserve accuracy under paraphrasing, outperforming accuracy-only prompt search while remaining computationally practical. The code is available for anonymous access at https://anonymous.4open.science/r/ATARE_TARE/.

Diffusion models have recently emerged as powerful priors for solving inverse problems. While computed tomography (CT) is theoretically a linear inverse problem, it poses many practical challenges. These include correlated noise, artifact structures, reliance on system geometry, and misaligned value ranges, which make the direct application of diffusion models more difficult than in domains like natural image generation. To systematically evaluate how diffusion models perform in this context and compare them with established reconstruction methods, we introduce DM4CT, a comprehensive benchmark for CT reconstruction. DM4CT includes datasets from both medical and industrial domains with sparse-view and noisy configurations. To explore the challenges of deploying diffusion models in practice, we additionally acquire a high-resolution CT dataset at a high-energy synchrotron facility and evaluate all methods under real experimental conditions. We benchmark nine recent diffusion-based methods alongside seven strong baselines, including model-based, unsupervised, and supervised approaches. Our analysis provides detailed insights into the behavior, strengths, and limitations of diffusion models for CT reconstruction. The real-world dataset is publicly available at zenodo.org/records/15420527, and the codebase is open-sourced at github.com/DM4CT/DM4CT.

Despite recent advances in state space models (SSMs) such as Mamba across various sequence domains, research on their standalone capacity for time series classification (TSC) has remained limited. We propose MambaSL, a framework that minimally redesigns the selective SSM and projection layers of a single-layer Mamba, guided by four TSC-specific hypotheses. To address benchmarking limitations—restricted configurations, partial University of East Anglia (UEA) dataset coverage, and insufficiently reproducible setups—we re-evaluate 20 strong baselines across all 30 UEA datasets under a unified protocol. Our results show that MambaSL achieves state-of-the-art performance on the UEA benchmark among 21 models, with statistically significant average improvements over baselines while ensuring reproducibility via public checkpoints.

The increasing deployment of powerful Multimodal Large Language Models (MLLMs), typically hosted on cloud platforms, urgently requires effective compression techniques to efficiently transmit signal inputs (e.g., images, videos) from edge devices with minimal bandwidth usage. However, conventional image codecs are optimized for fidelity to serve the Human Visual System (HVS) and ill-suited for MLLMs, in which diverse downstream tasks are jointly considered. In this paper, we first systematically analyze the impact of compression artifacts on several mainstream MLLMs. We find that: Compression distortion unevenly impacts different-level image features, leading to varying effects on MLLMs' downstream tasks depending on their feature-level reliance. Motivated by this discovery, we propose an image Codec TAilored to MLLMs (CoTAM) designed to adaptively protect multi-level features and suit different demands of downstream tasks. The encoder leverages CLIP's shallow-layer attention to generate an importance map for bit allocation, preserving critical semantic regions. Concurrently, the decoder integrates a lightweight adapter with a multi-level loss function to ensure the faithful reconstruction both of low-level details and high-level semantic context for robust synthesis of cross-level features. Extensive experiments validate that our method achieves up to 35.99\% bitrate saving while maintaining the same performance on the MLLM tasks, outperforming previous SOTA neural codecs.

Search agents have achieved significant advancements in enabling intelligent information retrieval and decision-making within interactive environments. Although reinforcement learning has been employed to train agentic models capable of more dynamic interactive retrieval, existing methods are limited by shallow tool-use depth and the accumulation of errors over multiple iterative interactions. In this paper, we present WebSeer, a more intelligent search agent trained via reinforcement learning enhanced with a self-reflection mechanism. Specifically, we construct a large dataset annotated with reflection patterns and design a two-stage training framework that unifies cold start and reinforcement learning within the self-reflection paradigm for real-world web-based environments, which enables the model to generate longer and more reflective tool-use trajectories. Our approach substantially extends tool-use chains and improves answer accuracy. Using a single 14B model, we achieve state-of-the-art results on HotpotQA and SimpleQA, with accuracies of 72.3\% and 90.0\%, respectively, and demonstrate strong generalization to out-of-distribution datasets.

Reinforcement learning (RL) training of large language models (LLMs) is limited by the policy's ability to generate rollouts with non-zero rewards: without such rewards, the policy is not updated and learning is stalled on hard problems, which are problems that the policy consistently fails to sample any correct rollouts for. We find that many hard problems remain unsolved due to the repeated generation of incorrect intermediate steps in a long reasoning trace; identifying and fixing these requires performing better \emph{credit assignment}. But existing approaches for credit assignment are either impractical or impose a substantial data-writing burden on oracles (\textit{e.g.}, humans). In this paper, we introduce \textbf{Interventional Training} (InT), a framework that leverages single-step oracle interventions to improve LLM reasoning. Given a reasoning attempt and ground-truth answer, the oracle detects and then provides language feedback on a single intermediate reasoning step, which is much cheaper than obtaining a full reasoning trace. \methodname{} then \emph{patches} the LLM by running supervised fine-tuning on the on-policy rollout up to the error, followed by the correction from the oracle. RL on this patched model now generates counterfactual traces and with merely $\approx$$100$ interventions from the oracle, \methodname{} solves 16\% more hard test problems that were previously unsolved (only zero rewards) and also improves performance across multiple standard evals.

Neural combinatorial optimization (NCO) has made significant advances in applying deep learning techniques to efficiently and effectively solve single-objective flexible job shop scheduling problems (FJSPs). However, the more practical multi-objective FJSPs (MOFJSPs) remain underexplored, limiting the applicability of NCO in multi-criteria decision-making scenarios. In this paper, we propose a decomposition-based NCO method to solve MOFJSPs. We present the dual conditional attention network (DCAN), a neural network architecture that takes the objective preferences along with the problem instance, aiming to learn adaptable policies over the preferences. By decomposing an MOFJSP into a set of subproblems with different preferences, the learned DCAN policies generate a set of solutions that reflect the corresponding trade-offs. We customize the Proximal Policy Optimization algorithm based on decomposition to effectively train the policy network for multiple objectives and define the state and reward based on combinations of different objectives. Extensive results showcase that our approach outperforms traditional multi-objective optimization methods and generalizes well across diverse types of problem instances.

Protein design stands as one of biology’s most important frontiers, with the potential to transform medicine, advance human health, and drive sustainability. Protein generation, a central task in protein design, has been greatly accelerated by AI-driven models—such as FoldFlow, MultiFlow, and AlphaFlow that build on residue-wise rigidity–based modeling pioneered by AlphaFold2. Residue-wise rigid-body representations reduce structural dimensionality while enforcing chemical constraints, enabling more efficient and physically consistent protein structure generation than all-atom modeling. Despite these advances, existing models often underutilize the vast structural information available in large-scale protein datasets. This highlights the importance of pretraining, which can provide richer representations and improve generalization across diverse protein design tasks. More importantly, the challenge lies in how to fully exploit abundant, low-cost unlabeled protein datasets using unsupervised pretraining. We introduce RigidSSL, a rigidity-based pretraining framework for proteins. RigidSSL canonicalizes structures into an inertial frame, employs a two-phase workflow combining large-scale perturbations and molecular dynamics views, and applies a rigid-body flow matching objective with Invariant Point Attention to capture global geometry. This enables learning stable, geometry-aware representations that improve downstream protein generation. To evaluate the effectiveness of RigidSSL, we conduct quantitative experiments on the protein generation task. Empirically, RigidSSL outperforms previous state-of-the-art geometric pretraining algorithms, leading to improvements in unconditional generation across all metrics, including designability, novelty, and diversity, for length up to 800 residues.

Test-time scaling improves the reasoning capabilities of large language models (LLMs) by allocating extra compute to generate longer Chains-of-Thoughts (CoTs). This enables models to tackle more complex problem by breaking them down into additional steps, backtracking, and correcting mistakes. Despite its strong performance--demonstrated by OpenAI's o1 and DeepSeek R1, the conditions in the training data under which long CoTs emerge, and when such long CoTs improve the performance, remain unclear. In this paper, we study the performance of test-time scaling for transformers trained on an in-context weight prediction task for linear regression. Our analysis provides a theoretical explanation for several intriguing observations: First, at any fixed test error, increasing test-time compute allows us to reduce the number of in-context examples (context length) in training prompts. Second, if the skills required to solve a downstream task are not sufficiently present in the training data, increasing test-time compute can harm performance. Finally, we characterize task hardness via the smallest eigenvalue of its feature covariance matrix and show that training on a diverse, relevant, and hard set of tasks results in best performance for test-time scaling. We confirm our findings with experiments on large, nonlinear transformer architectures.

In-context learning (ICL) has demonstrated significant performance improvements in transformer-based large models. This study identifies two key factors influencing ICL generalisation under complex non-i.i.d. scenario: algorithmic stability and distributional discrepancy. First, we establish a stability bound for transformer-based models trained with mini-batch gradient descent, revealing how specific optimization configurations interact with the smoothness of the loss landscape to ensure the stability of non-linear Transformers. Next, we introduce a distribution-level discrepancy measure that highlights the importance of aligning the ICL prompt distribution with the training data distribution to achieve effective generalisation. Building on these insights, we derive a generalisation error bound for ICL with asymptotic convergence guarantees, which further reveals that token-wise prediction errors accumulate over time and even lead to generalisation collapse if the prediction length is not properly constrained. Finally, empirical evaluations are provided to validate our theoretical findings.

LLM-based reasoning models have enabled the development of agentic systems that act as co-scientists, assisting in multi-step scientific analysis. However, evaluating these systems is challenging, as it requires realistic, end-to-end research scenarios that integrate data analysis, interpretation, and the generation of new insights from the experimental data. To address this limitation, we introduce HeurekaBench, a framework to create benchmarks with *exploratory, open-ended research questions* for experimental datasets. Each such question is grounded in a scientific study and its corresponding code repository, and is created using a semi-automated pipeline that leverages multiple LLMs to extract insights and generate candidate workflows, which are then verified against reported findings. We instantiate the framework in single-cell biology to obtain sc-HeurekaBench benchmark and use it to compare state-of-the-art single-cell agents. We further showcase the benefits of our benchmark for quantitatively analyzing current design choices in agentic systems. We find that the addition of a *critic* module can improve ill-formed responses for open-source LLM-based agents by up to 22% and close the gap with their closed-source counterparts. Overall, HeurekaBench sets a path toward rigorous, end-to-end evaluation of scientific agents, grounding benchmark construction in real scientific workflows.