ICML 2026 论文列表

The increasing situational awareness of language models raises safety concerns: models might be aware when they are evaluated, and adjust their behavior to evade monitoring and resist modification, e.g., pretending to be aligned only in evaluation. This *alignment faking* behavior is often interpreted as scheming: an intentional effort of strategic deception. In this paper, we examine an alternative interpretation, *performative misalignment*, which explains the change in behavior as a result of *sycophancy towards AI researchers*. To examine this hypothesis, we present three empirical findings. First, we show that evaluation awareness persists even when we tell models they are deployed, which contradicts the scheming story which predicts less misalignment when the model perceives evaluation. Second, we use probing and steering to show that our current methods cannot mechanistically distinguish sycophancy and scheming in alignment faking evaluations. Third, we fine-tune models to be more sycophantic and observe increased sensitivity to evaluation cues. To conclude, we emphasize deconfounding sycophancy from scheming for future work on evaluations and mitigations of intent misalignment.

Large language models are increasingly deployed in multi-turn interactive settings where users or environments can iteratively provide lightweight feedback. Unfortunately, optimizing such behavior presents a sharp dilemma in practice: online reinforcement learning is able to effectively address multi-turn dynamics but is prohibitively expensive due to the cost of generating full correction trajectories at every update, whereas offline supervised fine-tuning (SFT) is efficient but suffers from distribution shift and behavioral collapse. To this end, we novelly propose DRIFT (Decoupled Rollouts and Importance-Weighted Fine-Tuning), a framework that operationalizes the theoretical insight that the KL-regularized RL objective is equivalent to importance-weighted supervised learning. DRIFT decouples rollout from optimization by sampling offline interaction trajectories from a fixed reference policy, deriving return-based importance weights, and optimizing the policy via weighted SFT on the resulting dataset. Empirically, we demonstrate that DRIFT matches or exceeds the performance of multi-turn reinforcement learning baselines while maintaining the training efficiency and simplicity of standard supervised fine-tuning. Code is available at https://github.com/2020-qqtcg/DRIFT.

Large Reasoning Models (LRMs) achieve remarkable inference-time improvements through parallel thinking. However, existing methods rely on redundant sampling of reasoning trajectories, failing to effectively explore the reasoning space to uncover high-quality solutions. To address these limitations, we propose **D**ecoding **T**ree **S**ketching (DTS), a plug-and-play decoding framework for structural multi-trajectory exploration and reasoning selection. For reasoning exploration, DTS sketches a backbone tree of the reasoning space by selectively branching at decision tokens. For reasoning selection, guided by length-accuracy anti-correlation, DTS designs an early termination to prioritize short and reliable trajectories during decoding. Experimental results across four LRMs and datasets demonstrate that DTS significantly enhances accuracy by **14\%** and reduces repetitive generation by **8\%** on average. Notably, DTS enables smaller models to outperform larger models with 10$\times$ the size, highlighting its potential to strengthen reasoning capabilities.

Textual Gradient-style optimizers (TextGrad) enable gradient-like feedback propagation through compound AI systems. However, they do not work well for deep chains. The root cause of this limitation stems from the *Semantic Entanglement* problem in these extended workflows. In standard textual backpropagation, feedback signals mix local critiques with upstream contexts, leading to *Attribution Ambiguity*. To address this challenge, we propose TextResNet, a framework that reformulates the optimization process to achieve precise signal routing via four key innovations. Firstly, in the forward pass, it enforces Additive Semantic Deltas to preserve an Identity Highway for gradient flow. Secondly, in the backward pass, it introduces Semantic Gradient Decomposition via a Semantic Projector to disentangle feedback into causally independent subspaces. Thirdly, it implements Causal Routing, which routes projected signals to their specific components. Finally, it performs Density-Aware Optimization Scheduling to leverage the disentangled signals to dynamically allocate resources to key system bottlenecks. Our results show that TextResNet not only achieves superior performance compared to TextGrad, but also exhibits remarkable stability for agentic tasks in compound AI systems where baselines collapse. Code is available at https://github.com/JeanDiable/TextResNet.

Large language models (LLMs) face a "loyalty dilemma" when correctness is conditioned on an active world-of-discourse. We identify a systemic failure mode---world misattribution---where models implicitly ground generation in an incompatible regime and drift from the target world. We propose World Entropy Tethering (WET), an inference-time monitor-and-tether: a world-entropy probe flags drift risk on prompt anchors, and a conditional score matching geometry model identifies tethering heads for entropy-gated rescaling. Experiments show: (I) Linear Separability: world labels are linearly decodable from internal states; (II) Geometric Drift: hallucinations are preceded by measurable deviations from the target world region; and (III) Targeted Mitigation: WET improves world consistency and reduces hallucination rates by up to 22.4% without compromising generation quality. Code is available at https://github.com/guess-guess-who-i-am/WET-World-Entropy-Tethering.

Test-time policy adaptation for multi-turn interactions (T$^2$PAM) is essential for aligning Large Language Models (LLMs) with dynamic user needs. However, existing paradigms typically treat adaptation as a single-axis problem by either purely refining instructions or solely updating weights. This bifurcated approach overlooks the fact that interaction failures arise from a coupled mixture of context ambiguity and model incapacity. To address this, we propose ROSA2, a framework that reformulates T$^2$PAM as a joint optimization problem over the heterogeneous space of Words and Weights. Within this framework, the semantic stream acts as a feedback normalizer that transforms noisy or ambiguous user feedback into actionable instructions, ensuring that parametric adaptation is performed on semantically clarified trajectories. Theoretically, we prove that this semantic pre-conditioning strictly reduces the required parameter shift for convergence. Empirically, ROSA2 consistently outperforms state-of-the-art baselines across diverse mathematical, general reasoning, and coding benchmarks. It achieves up to a 37.8\% accuracy improvement while reducing average interaction turns by 40\%, demonstrating that ROSA2 unlocks the true potential of parameter updates.

Reinforcement learning (RL) based post-training for explicit chain-of-thought (e.g., GRPO) improves the reasoning ability of multimodal large-scale reasoning models (MLRMs). But recent evidence shows that it can simultaneously degrade safety alignment and increase jailbreak success rates. We propose SafeThink, a lightweight inference-time defense that treats safety recovery as a satisficing constraint rather than a maximization objective. SafeThink monitors the evolving reasoning trace with a safety reward model and conditionally injects an optimized short corrective prefix ("Wait, think safely") only when the safety threshold is violated. In our evaluations across six open-source MLRMs and four jailbreak benchmarks (JailbreakV-28K, Hades, FigStep, and MM-SafetyBench), SafeThink reduces attack success rates by 30-60 % (e.g., LlamaV-o1: 63.33\% $\rightarrow$5.74\% on JailbreakV-28K, R1-OneVision: 69.07\%$\rightarrow$5.65\% on Hades) while preserving reasoning performance (MathVista accuracy: 65.20\%$\rightarrow$65.00\%). A key empirical finding from our experiments is that safety recovery is often only a few steering steps away: intervening in the first $1–3$ reasoning steps typically suffices to redirect the full generation toward safe completions.

Autoregressive models generate sequences monotonically, where any sampled token, even if erroneous or sub-optimal, becomes a permanent condition for all subsequent steps. This structural limitation means that autoregressive models cannot revisit or revise earlier decisions, i.e., a capability essential for complex generation tasks where exploration and correction are necessary. To this end, we propose N-MARS, a Non-Monotonic AutoregRessive Sequence modeling framework that enables models to generate, evaluate, and revise tokens within a single forward pass, effectively allowing exploration before commitment. We operationalize this framework through a learned erase token <UNDO> that retracts the previous token, enabling on-the-fly revision within standard autoregressive decoding. To train the model, we introduce a sequence augmentation method that constructs error-correction trajectories from model-generated deviations paired with ground-truth references. We then propose masked supervised fine-tuning (mSFT) that exposes the model to errors as context for learning when to revise, without optimizing their likelihood. Finally, we refine the model with group relative policy optimization (GRPO) that incentivizes judicious <UNDO> usage, i.e., rewarding effective corrections while penalizing unsuccessful ones. We conduct comprehensive theoretical and empirical analysis to validate the effectiveness of N-MARS, demonstrating a robust foundation for non-monotonic sequence modeling.

Large generative models raise growing concerns about provenance, misinformation, and impersonation. Digital watermarking offers a principled solution, yet extending it to natural language remains challenging due to text discreteness and sensitivity to semantic perturbations. Existing text watermarking methods either operate at the token level requiring white-box access and remaining fragile to paraphrasing, or at the sentence level, which supports black-box deployment but suffers from low Watermark Success Rate (WSR). We show that low WSR in sentence-level watermarking primarily stems from low injection success probability caused by a mismatch between posterior embedding distributions and semantic accept regions. Based on this insight, we propose \textbf{X-Guard}, a geometry-aware sentence-level watermarking framework that improves injection success by jointly optimizing embedding distributions and semantic space partitioning. X-Guard learns a more isotropic embedding space and introduces \textbf{A$^2$PQ}, a centroid-aligned partitioning scheme that approximately equalizes probability mass across regions. Extensive experiments across multiple models, languages, and attack settings demonstrate that X-Guard consistently improves robustness while preserving text fluency and practical deployability.

Large language models have advanced web agents, yet current agents lack personalization capabilities. Since users rarely specify every detail of their intent, practical web agents must be able to interpret ambiguous queries by inferring user preferences and contexts. To address this challenge, we present Persona2Web, the first benchmark for evaluating personalized web agents on the real open web, built upon the clarify-to-personalize principle, which requires agents to resolve ambiguity based on user history rather than relying on explicit instructions. Persona2Web consists of: (1) user histories that reveal preferences implicitly over long time spans, (2) ambiguous queries that require agents to infer implicit user preferences, and (3) a reasoning-aware evaluation framework that enables fine-grained assessment of personalization. We conduct extensive experiments across various agent architectures, backbone models, history access schemes, and queries with varying ambiguity levels, revealing key challenges in personalized web agent behavior. For reproducibility, our codes and datasets are publicly available at https://serin-kimm.github.io/Persona2Web/

Reinforcement learning significantly enhances LLM capabilities but suffers from a critical issue: length inflation, where models adopt verbosity or inefficient reasoning to maximize rewards. Prior approaches struggle to address this challenge in a general and lossless manner, primarily because additive penalties introduce a compensatory effect that creates optimization shortcuts, while heuristic gating strategies lack generality beyond binary feedback. To bridge this gap, we present Group Relative Reward Rescaling (GR$^3$), which reframes length control as a multiplicative rescaling paradigm, effectively establishing a generalized, continuous, and reward-dependent gating mechanism. To further ensure lossless optimization, we incorporate group-relative regularization and advantage-aware calibration, which dynamically adapt length budgets to instance difficulty and preserve the advantage signal of high-quality trajectories. Empirically, across both RLHF and RLVR settings, GR$^3$ maintains training dynamics and downstream performance comparable to standard GRPO while significantly mitigating length inflation, outperforming state-of-the-art length-regularized baselines.

LLM pre-training efficacy increasingly depends on data composition rather than sheer volume. Yet, optimal mixing is hindered by categorization flaws: human taxonomies suffer from ontological misalignment, and Euclidean clustering fails to address embedding anisotropy. We introduce **GEM** (**G**eometric **E**ntropy **M**ixing), a framework reformulating data curation as a variational problem on the hypersphere augmented with a **mixing-balance regularizer**. By decoupling the generative prior and optimizing the objective via a provable **MM (Minorize-Maximize)** algorithm, GEM effectively counteracts the cluster collapse to discover balanced semantic structures invisible to Euclidean heuristics. We employ teacher-student distillation to scale this geometric fidelity to web-scale corpora and introduce the **Geometric Influence Score (GIS)** for interpretable taxonomy generation. Experiments with 1.1B-parameter models demonstrate that GEM establishes a new state-of-the-art when integrated into mixing strategies like DoReMi and RegMix, improving average downstream accuracy by up to **1.2%** and offering a robust coordinate system for predictable data mixing.

Self-evolving memory systems are rapidly reshaping the evolutionary paradigm of large language model (LLM)-based agents. Prior work has predominantly relied on manually engineered memory architectures to store trajectories, distill experience, and synthesize reusable tools, enabling agents to evolve on the fly within environment interactions. However, this paradigm is fundamentally constrained by the \textit{staticity} of the memory system itself: while memory facilitates agent-level evolving, the underlying memory architecture cannot be meta-adapted to diverse task contexts. To address this gap, we propose MemEvolve, a meta-evolutionary framework that jointly evolves agents’ experiential knowledge and their memory architecture, allowing agent systems not only to accumulate experience but also to progressively refine how they learn from it. To ground MemEvolve in prior work and promote openness in future self-evolving systems, we introduce EvolveLab, a unified memory codebase that distills twelve representative memory systems into a modular design space (\textit{encode}, \textit{store}, \textit{retrieve}, \textit{manage}), providing a standardized implementation substrate and a fair experimental arena. Extensive evaluations on four challenging agentic benchmarks show that MemEvolve delivers (i) substantial performance gains, improving frameworks such as SmolAgent and Flash-Searcher by up to $17.06\%$, and (ii) strong cross-task and cross-LLM generalization, yielding memory architectures that transfer effectively across diverse benchmarks and backbones.

Despite recent advancements in Multi-modal Large Language Models (MLLMs) on diverse understanding tasks, these models struggle to solve problems which require extensive multi-step reasoning. This is primarily due to the progressive dilution of visual information during long-context generation, which hinders their ability to fully exploit test-time scaling. To address this issue, we introduce Vision-aligned Latent Reasoning (VaLR), a simple, yet effective reasoning framework that dynamically generates vision-aligned latent tokens before each Chain of Thought reasoning step, guiding the model to reason based on perceptual cues in the latent space. Specifically, VaLR is trained to preserve visual knowledge during reasoning by aligning intermediate embeddings of MLLM with those from vision encoders. Empirical results demonstrate that VaLR consistently outperforms existing approaches across a wide range of benchmarks requiring long-context understanding or precise visual perception, while exhibiting test-time scaling behavior not observed in prior MLLMs. In particular, VaLR improves the performance significantly from 33.0\% to 52.9\% on VSI-Bench, achieving a 19.9\%p gain over Qwen2.5-VL.

Large language models (LLMs) are enabling reasoning over 2D and 3D structures, yet existing methods remain modality-specific and typically compress structural inputs through sequence-based tokenization or fixed-length query connectors. Such architectures either omit the geometric grounding requisite for mitigating structural hallucinations, or impose inflexible modality fusion bottlenecks that concurrently over-compress and suboptimally allocate structural tokens, thereby impeding the realization of generalized all-atom reasoning. We introduce **Cuttlefish**, a unified multimodal LLM that grounds language reasoning in geometric cues while scaling modality tokens with structural complexity. First, **Scaling-Aware Patching** leverages an instruction-conditioned gating mechanism to generate variable-size patches over structural graphs, adaptively scaling the query token budget with structural complexity to mitigate fixed-length connector bottlenecks. Second, **Geometry Grounding Adapter** refines these adaptive tokens via cross-attention to modality embeddings and injects the resulting modality tokens into the LLM, exposing explicit geometric cues to reduce structural hallucination. Experiments across interdisciplinary all-atom benchmarks demonstrate that Cuttlefish achieves superior performance in heterogeneous structure-grounded reasoning. Code: github.com/zihao-jing/Cuttlefish.

Continual incorporation of new knowledge is essential for the long-term evolution of large language models (LLMs). Existing approaches typically rely on parameter-update algorithms to mitigate catastrophic forgetting, yet they suffer from fundamental limitations: 1) forgetting is unavoidable as the amount of newly injected knowledge grows; and 2) model updates are often irreversible. As modern LLMs become increasingly expressive, it is natural to question whether large-scale weight updates are necessary for acquiring a small amount of new knowledge. In this work, we propose a principled framework that models autoregressive language generation as a Markov process over tokens, where model memory is represented by a Markov transition matrix. Under this formulation, incorporating new knowledge/tokens corresponds to extending the state space, and preserving existing transitions guarantees retention of previously learned knowledge. We then prove a sample complexity bound for incorporating new tokens via a token-to-dictionary mapping strategy. In particular, for learning the transition behavior of each new token, the required number of samples scales linearly with the number of existing tokens it is mapped to. To realize this mapping, we propose an embedding-tuning algorithm that requires minimal parameter updates and induces zero forgetting. Experimental results further demonstrate the effectiveness of our method and validate our theoretical findings.

Weight-only quantization has become a standard approach for efficiently serving large language models (LLMs). However, existing methods fail to efficiently compress models to binary (1-bit) levels, as they either require large amounts of data and compute or incur additional storage. In this work, we propose NanoQuant, a post-training quantization (PTQ) method to compress LLMs to both binary and sub-1-bit levels. NanoQuant formulates quantization as a low-rank binary factorization problem, and compresses full-precision weights to low-rank binary matrices and scales. Specifically, it utilizes an efficient alternating direction method of multipliers (ADMM) solver to precisely initialize latent binary matrices and scales, and then tunes the initialized parameters through a block and model reconstruction process. Consequently, NanoQuant establishes a new Pareto frontier in low-memory post-training quantization, and enables sub-1-bit compression. NanoQuant makes large-scale deployment feasible on consumer hardware. For example, it compresses Llama-2-70B by 24$\times$ in just 13 hours on a single H100, enabling a 70B model to operate on a consumer 8 GB GPU.

Large Reasoning Models (LRMs) have achieved remarkable progress thanks to Reinforcement Learning with Verifiable Rewards (RLVR) on Chain-of-Thoughts (CoTs). However, since long CoTs naturally contain trial and errors and mainstream RLVR approaches choose outcome-correct CoT trajectories for memorization, the redundant explorations in long CoTs are inevitably reinforced, which results in the over-thinking issues of LRMs. Previous attempts to resolve this issue mainly give more advantage to shorter trajectories, yet their learning signals are still outcome-based and cannot reduce the memorization of redundant explorations in long CoTs. Therefore, we propose ThoughtFold, a framework that leverages fine-grained preference learning to mitigate redundant explorations for efficient reasoning. ThoughtFold employs an introspective strategy to identify redundancy within each correct trajectory, which yields a spectrum of candidate sub-trajectories. Leveraging this spectrum, we introduce a masked preference optimization objective that explicitly penalizes redundant explorations and encourages the model to directly bridge essential reasoning segments, effectively folding its reasoning chains into a more concise path. Extensive experiments show that ThoughtFold significantly enhances efficiency. It reduces the token usage of DeepSeek-R1-Distill-Qwen-7B by approximately 56\% while maintaining state-of-the-art accuracy.

Autonomous agents have recently achieved remarkable progress across diverse domains, yet most evaluations focus on short-horizon, fully observable tasks. In contrast, many critical real-world tasks, such as large-scale software development, commercial investment, and scientific discovery, unfold in long-horizon and partially observable scenarios where success hinges on sustained reasoning, planning, memory management, and tool use. Existing benchmarks rarely capture these long-horizon challenges, leaving a gap in systematic evaluation. To bridge this gap, we introduce $\textbf{UltraHorizon}$, a novel benchmark that measures the foundational capabilities essential for complex real-world challenges. We use exploration as a unifying task across three distinct environments to validate these core competencies. Agents are designed in long-horizon discovery tasks where they must iteratively uncover hidden rules through sustained reasoning, planning, memory and tools management, and interaction with environments. Under the heaviest scale setting, trajectories average $\textbf{200k+}$ tokens and $\textbf{400+}$ tool calls, whereas in standard configurations they still exceed $\textbf{35k}$ tokens and involve more than $\textbf{60}$ tool calls on average. Our extensive experiments reveal that agents powered by state-of-the-art LLMs consistently underperform in these settings, whereas human participants achieve much higher scores, underscoring a persistent gap in agents' long-horizon exploration abilities. We also observe that simple scaling fails in our task. To better illustrate the failure of agents, we conduct an in-depth analysis of collected trajectories. We identify eight types of errors and attribute them to two primary causes: in-context locking and functional fundamental capability gaps.