ICML 2026 Papers

Recent advances in large language models (LLMs) have produced many specialized multimodal LLMs (MLLMs) that share common foundational LLMs, forming distinct model lineages. It remains unclear whether a fundamental behavioral link exists between the foundational LLMs and downstream variants. We investigate this question by quantifying head-level context-truthfulness scores. Across diverse LLM and MLLM lineages, including Vicuna-, Qwen2.5-, LLaMA2-, and Mistral-based models, we find that Truth Scores are strongly preserved within model families, even after instruction tuning or multimodal adaptation. We further show that this inheritance is consistent with attention-head weight preservation, and that context-truthful heads attend to query-relevant evidence. Building on this finding, we propose TruthProbe, a soft-gating strategy that amplifies context-truthful heads while preserving other head contributions. TruthProbe improves contextual truthfulness on HaluEval and reduces multimodal hallucination on POPE and CHAIR, with base-LLM Truth Scores transferring effectively to their fine-tuned LLM and MLLM descendants. Code is available at https://github.com/miso-choi/TruthProbe.

Discrete diffusion language models generate text by iteratively denoising an entire response in parallel. At each step, they predict tentative tokens for every masked position, committing the confident predictions to the output and discarding the unconfident ones. We show that the discarded tokens are in fact a useful lookahead signal for retrieval-augmented generation: even low-confidence tokens often surface salient entities early in the denoising trajectory, enabling retrieval of stronger evidence before the output is finalized. We exploit this through Self-Augmenting Retrieval for Diffusion Language Models (SARDI), a dynamic RAG framework that uses these lookahead tokens to guide retrieval during denoising. SARDI is training-free, retriever-agnostic, and applicable to any reasoning-capable discrete diffusion language model. Across five multi-hop QA benchmarks, SARDI outperforms current training-free diffusion and autoregressive retrieval baselines at up to $8\times$ higher throughput.

Training large language models (LLMs) is computationally expensive, partly because the loss exhibits slow power-law convergence whose origin remains debatable. Through systematic analysis of toy models and empirical evaluation of LLMs, we show that this behavior can arise intrinsically from the use of softmax and cross-entropy. When learning peaked probability distributions, e.g., next-token distributions, these components generically yield power-law vanishing losses and gradients, regardless of many microscopic details, creating a fundamental optimization bottleneck. This ultimately leads to power-law time scaling of the loss with a universal exponent of $1/3$. Our results provide a mechanistic explanation for observed neural scaling and suggest new directions for improving LLM training efficiency.

Gradient signals in LLM training are highly anisotropic: recurrent linguistic structure concentrates energy into a small set of dominant spectral directions, while context-specific information resides in a long tail. We show that this spike–tail separation persists throughout training, with the spike occupying only about 1.5% of directions yet dominating optimizer statistics. This dominance suppresses tail learning by contracting tail updates through second-moment normalization and tightening the globally stable learning-rate bound. Motivated by this analysis, we propose Spectra, a spike-aware optimizer that suppresses the dominant low-rank spike subspace without amplifying the noise-sensitive spectral tail. Spectra tracks the spike subspace via cached, warm-started power iteration and applies low-rank spectral shaping with negligible overhead and substantially reduced optimizer-state memory. Across Qwen3-0.6B trained on 100B tokens and LLaMA3-8B trained on 50B tokens, Spectra achieves the lowest final validation loss, improving average downstream accuracy by 1.41/0.89 and 1.62/0.66 points over AdamW/Muon, respectively. For wall-clock convergence, Spectra reaches matched loss targets up to 1.31×, 1.34×, and 1.24× faster than AdamW on Qwen3-0.6B, Qwen3-2B-A0.8B, and Qwen3-8B; its speedup over Muon grows as model scale increases from 0.6B to 8B. For computational efficiency, Spectra is 5.1× faster than Muon in optimizer processing time, cuts optimizer-state memory by 49.25%, and achieves the lowest measured end-to-end per-step runtime. Spectra's Megatron integration is available at https://github.com/kimmichtank/spectra.

LLM parameter editing methods commonly rely on computing an ideal target hidden-state at a target layer (referred as anchor point) and distributing the target vector to multiple preceding layers (commonly known as backward spreading) for cooperative editing. Although widely used for a long time, its underlying basis have not been systematically investigated. In this paper, we first conduct a systematic study of its foundations, which helps clarify its capability boundaries, practical considerations, and potential failure modes. Then, we propose a simple and elegant alternative that replaces backward spreading with forward-propagation. Instead of optimizing the target at the last editing layer, we optimize the anchor point at the first editing layer, and then propagate it forward to obtain accurate and mutually compatible target hidden-states for all subsequent editing layers. This approach achieves the same computational complexity as existing methods while producing more accurate layer-wise targets. Our method is simple, without interfering with either the computation of the initial target hidden state or any other components of the subsequent editing pipeline, and thus constituting a benefit for a wide range of LLM parameter editing methods.

Loss spikes remain a persistent obstacle in large-scale language model pretraining. While previous research has attempted to identify the root cause of loss spikes by investigating individual factors, we observe that, in practice, such spikes are typically triggered by the confluence of heterogeneous factors. Empirically, loss spikes may arise from a combination of data outliers, hardware or transient computational faults, numerical precision issues, and hyperparameter settings. Regardless of the underlying cause, these spikes manifest as unstable optimizer updates, as abnormal gradients contaminate both first- and second-moment states. In this paper, we propose a principled gradient-centric remedy: AdaGC, an adaptive per-tensor gradient clipping scheme that mitigates such contamination by bounding gradient norms relative to a tensor-wise exponential moving average of their historical clipped values. AdaGC is optimizer-agnostic, introduces negligible memory overhead, and reduces communication costs compared to GlobalGC, particularly in hybrid-parallel distributed training. Experiments on Llama-2 7B, Mixtral 8×1B, and ERNIE 10B-A1.4B demonstrate that AdaGC robustly eliminates training instabilities, consistently reducing spike scores to zero for all models and improving downstream accuracy over GlobalGC by 1.32\%, 1.27\%, and 2.48\%, respectively. Furthermore, AdaGC seamlessly integrates with optimizers such as Muon and Lion, consistently yielding higher average accuracy and zero spike scores. The code is available at https://github.com/PaddlePaddle/PaddleFleet (see Research/AdaGC).

Effective tool use and reasoning are essential capabilities for large reasoning models (LRMs) to address complex real-world problems. Through empirical analysis, we identify that current LRMs lack the capability of sub-task decomposition in complex tool use scenarios, leading to Lazy Reasoning. To address this, we propose a two-stage training framework D-CORE (Decomposing tasks and Composing Reasoning processes) that first incentivize the LRMs' task decomposition reasoning capability via self-distillation, followed by diversity-aware reinforcement learning (RL) to restore LRMs' reflective reasoning capability. D-CORE achieves robust tool-use improvements across diverse benchmarks and model scales. Experiments on BFCLv3 demonstrate superiority of our method: D-CORE-8B reaches 77.7\% accuracy, surpassing the best-performing 8B model by 5.7\%. Meanwhile, D-CORE-14B establishes a new state-of-the-art at 79.3\%. Code is available at https://github.com/alibaba/EfficientAI; the model is available at https://huggingface.co/bowiehsu/D-CORE-8B; and the data are available at https://huggingface.co/datasets/bowiehsu/D-CORE-Self-Distillation-50K.

Large language models are increasingly evaluated as interactive agents, yet standard agent benchmarks conflate two qualitatively distinct sources of success: semantic tool-use and interface-specific interaction pattern memorization. Because both mechanisms can yield identical task success on the original interface, benchmark scores alone are not identifiable evidence of environment-invariant capability. We propose **PIPE**, a protocol-level evaluation augmentation for diagnosing interface reliance by minimally rewriting environment interfaces while preserving task semantics and execution behavior. Across 16 environments from AgentBench and AgentGym and a range of open-source and API-based agents, PIPE reveals that task-specific trajectory-SFT can amplify reliance on training-time interface forms: in several environments, agents with trajectory-SFT degrade sharply under minimal interface rewrites, whereas other agents are often more stable. We further introduce Interface Reliance (IR), a counterbalanced alias-based metric that quantifies preference for training-time interfaces, and show that interface shortcutting exhibits environment-dependent, non-monotonic training dynamics that remain invisible under standard evaluation. Our code is available at https://github.com/ChengZe2005/What-Do-Agents-Learn-from-Trajectory-SFT-Semantics-or-Interfaces-.

Diffusion Language Models (DLMs) have emerged as a flexible alternative to autoregressive (AR) models. They can decode tokens in any order, but the generation quality critically depends on the decoding strategy. Existing approaches predominantly rely on local heuristics, such as confidence or entropy, which may fail to capture sequence-level dependencies and the semantics in the context. To solve this problem, we propose Latent-aware Unmasking Guidance Search (LUGS), a novel decoding framework that leverages the model's internal hidden states to guide the unmasking process. By incorporating latent-aware scores to compensate for the limitations of local heuristics such as confidence or entropy, LUGS improves the model's performance. Extensive experiments on various downstream tasks demonstrate that our approach consistently outperforms existing baselines on LLaDA-8B-instruct and LLaDA-1.5 models. In Science and Reason tasks, LUGS improved performance by more than 1\% on both base models. And LUGS obtains an average improvement of 3.5\% in code generation. Remarkably, LUGS outperforms the beam search baseline by more than 5\% on average using LLaDA-8B-Instruct on code tasks. These results highlight the potential of latent-aware guidance for advancing controllable and high-quality generation.

Large reasoning models achieve strong performance by scaling inference-time chain-of-thought, but this paradigm suffers from quadratic cost, context length limits, and degraded reasoning due to lost-in-the-middle effects. Iterative reasoning mitigates these issues by periodically summarizing intermediate thoughts, yet existing methods rely on supervised learning or fixed heuristics and fail to optimize when to summarize, what to preserve, and how to resume reasoning. We propose InftyThink+, an end-to-end reinforcement learning framework that optimizes the entire iterative reasoning trajectory, building on model-controlled iteration boundaries and explicit summarization. InftyThink+ adopts a two-stage training scheme with supervised cold-start followed by trajectory-level reinforcement learning, enabling the model to learn strategic summarization and continuation decisions. Experiments on DeepSeek-R1-Distill-Qwen-1.5B show that InftyThink+ improves accuracy by 21% on AIME24 and outperforms conventional long chain-of-thought reinforcement learning by a clear margin, while also generalizing better to out-of-distribution benchmarks. Moreover, InftyThink+ significantly reduces inference latency and accelerates reinforcement learning training, demonstrating improved reasoning efficiency alongside stronger performance.

We introduce **Native Parallel Reasoner (NPR)**, a teacher-free framework that enables Large Language Models (LLMs) to self-evolve genuine parallel reasoning capabilities. NPR transforms the model from sequential emulation to native parallel cognition through three key innovations: (1) a **self-distilled** progressive training paradigm that transitions from cold-start format discovery to strict topological constraints without external supervision; (2) a novel **Parallel-Aware Policy Optimization (PAPO)** algorithm that optimizes branching policies directly within the execution graph, allowing the model to learn adaptive decomposition via trial and error; and (3) a robust **NPR Engine** that refactors memory management and flow control of SGLang to enable stable, large-scale parallel RL training. Across eight reasoning benchmarks, NPR trained on Qwen3-4B achieves performance gains of up to 24.5\% and inference speedups up to 4.6$\times$. Unlike prior baselines that often fall back to autoregressive decoding, NPR demonstrates 100\% genuine parallel execution, establishing a new standard for self-evolving, efficient, and scalable agentic reasoning.

A widely adopted strategy for model enhancement is to use synthetic data generated by a stronger model for supervised fine-tuning (SFT). However, for emerging reasoning models like Qwen3-8B, this approach often fails to improve reasoning capabilities and can even lead to a substantial drop in performance. In this work, we identify substantial stylistic divergence between teacher generated data and the distribution of student as a major factor impacting SFT. To bridge this gap, we propose a Teacher–Student Cooperation Data Synthesis framework (TESSY), which interleaves teacher and student models to alternately generate style and non-style tokens. Consequently, TESSY produces synthetic sequences that inherit the advanced reasoning capabilities of the teacher while maintaining stylistic consistency with the distribution of the student. In experiments on code generation using GPT-OSS-120B as the teacher, fine-tuning Qwen3-8B on teacher-generated data leads to performance drops of 3.25% on LiveCodeBench-Pro and 10.02% on OJBench, whereas TESSY achieves improvements of 11.25% and 6.68%.

Large language models (LLMs) handle many tasks with one set of parameters, but under KV-cached inference it is unclear what task-general structure, if any, is used at $\textit{decode time}$ rather than during $\textit{prefill}$. We propose $\textbf{DecodeShare}$, a protocol that identifies a low-dimensional subspace consistently shared across tasks in decode-time hidden states, and then tests its causal role by removing that subspace only during decoding. In our experiments, disturbing the discovered shared subspace degrades decision performance far more than disturbing either a prefill-derived or random subspace under the same intervention budget. We further show this decode-shared subspace has practical consequences for activation steering: common steering directions can overlap the task-general decode channel. Projecting out this shared subspace directly separates the functional roles of the two components, while evaluating steering vectors at decode-time yields more reliable signal for downstream deployment than prefill-based proxies. Despite its compactness, the shared subspace can serve as a high-leverage causal channel at decode time. Code is available at: https://github.com/Zishan-Shao/decodeshare.git.

Recent work has identified a counterintuitive phenomenon termed “Hyperfitting", where fine-tuning Large Language Models (LLMs) to near-zero training loss on small datasets surprisingly enhances open-ended generation quality and mitigates repetition in greedy decoding. While effective, the underlying mechanism remains poorly understood, with the extremely low-entropy output distributions suggesting a potential equivalence to simple temperature scaling. In this work, we demonstrate that this phenomenon is fundamentally distinct from distribution sharpening; entropy-matched control experiments reveal that temperature scaling fails to replicate the diversity gains of hyperfitting. Furthermore, we falsify the hypothesis of static vocabulary reweighting, showing through ablation studies that hyperfitting relies on a dynamic, context-dependent rank reordering mechanism. Layer-wise analysis localizes this effect to a “Terminal Expansion" in the final transformer block, where a substantial geometric expansion of the feature space ($\Delta \mathrm{Dim} \approx +80.8$) facilitates the promotion of deep-tail tokens. Additionally, we introduce \textbf{Late-Stage LoRA}, a targeted fine-tuning strategy that updates only the final 5 layers, yielding robust generation with minimal parameter updates.

Large reasoning models achieve strong performance through test-time scaling, but this incurs substantial computational overhead due to long decoding from short prompts. While sparse attention can reduce latency and memory usage, existing methods often degrade reasoning accuracy because selection errors accumulate over long generation horizons, or require costly retraining. We introduce LessIsMore, a training-free sparse attention mechanism for long-horizon reasoning. Our key insight is that token importance in reasoning is global and stable: critical tokens are largely shared across attention heads and remain stable over decoding steps. Guided by this structure, LessIsMore enforces cross-head unified token selection and preserves recent context via a stable recency window, yielding a globally consistent token set that can be reused across layers. Across multiple model families and challenging reasoning benchmarks, LessIsMore matches or improves accuracy while attending to substantially fewer tokens. With kernel-level optimizations, LessIsMore achieves up to $1.6\times$ end-to-end decoding speedup and up to $1.72\times$ faster sparse attention computation, with additional long-context results demonstrating the generality of our approach.

Large language models (LLMs) have shown promising potential in materials science, enabling tasks ranging from knowledge retrieval to property prediction. Existing materials science benchmarks mainly focus on perceptual or knowledge-based tasks, largely ignoring the structure modelling tasks, a core challenge in real scientific workflows. In practice, constructing and manipulating atomic structures is one of the most creative and least automated steps in materials research. In this work, we introduce AtomWorld, a benchmark designed to evaluate the abilities of LLMs on structure modifications. The benchmark includes ten fundamental actions under four widely used modelling categories, enabling verifiable evaluation metrics. We find that Gemini 2.5 Pro generally performs the best. While the success rate decreases markedly with increasing modelling complexity, with particularly low success rates (below 12\% for rotation) for operations involving complex spatial relations. Our results suggest that contemporary LLMs are better suited as copilots for materials structure modelling rather than fully unsupervised autonomous scientific agents. Beyond evaluation, AtomWorld also serves as a testbed and playground for developing future structure-aware models, including reinforcement learning and agentic approaches.

We show that reinforcement learning with verifiable rewards (RLVR) can elicit strong mathematical reasoning in certain language models even with spurious rewards that have little, no, or outright negative correlation with the correct answer. For example, RLVR training with GRPO improves MATH-500 performance for Qwen2.5-Math-7B in absolute points by 21.4% using randomly assigned rewards, nearly matching the 29.1% gained with ground truth rewards. To explain this counterintuitive observation, we show that GRPO exhibits a clipping bias arising from the clip term, which can amplify high-prior behaviors learned during pre-training even without informative rewards. As a case study, we identify one such high-prior behavior for Qwen2.5-Math models, which we term code reasoning---reasoning in code without actual code execution; code reasoning frequency increases from 65% to over 90% with spurious rewards. However, the presence of such amplifiable behaviors is highly model-dependent. In practice, spurious rewards that are effective for Qwen models often fail to produce gains for other model families, such as Llama3 or OLMo2. Our results highlight the importance of validating RL methods across diverse models rather than relying on a single de facto choice: large performance gains can arise on Qwen models even from random rewards that do not reflect genuine capability improvements.

Large Language Models (LLMs) exhibit significant safety disparities across languages, with low-resource languages (LRLs) often bypassing safety guardrails established for high-resource languages (HRLs) like English. Existing solutions, such as multilingual supervised fine-tuning (SFT) or Reinforcement Learning from Human Feedback (RLHF), are computationally expensive and de- pendent on scarce multilingual safety data. In this paper, we propose a novel, training-free alignment framework based on Sparse Weight Editing. Identifying that safety capabilities are localized within a sparse set of ”safety neurons,” we formulate the cross-lingual alignment problem as a constrained linear transformation. We derive a closed-form solution to optimally map the harmful representations of LRLs to the robust safety subspaces of HRLs, while preserving general utility via a null-space projection constraint. Extensive experiments across 8 languages and multiple model families (Llama-3, Qwen-2.5) demonstrate that our method significantly reduces Attack Success Rate (ASR) in LRLs with negligible impact on general reasoning capabilities, all achieved with a single, data-efficient calculation.

Large language models (LLMs) have achieved remarkable breakthroughs across various applications. However, their architectures remain inefficient in pretraining due to two main limitations: (i) self-attention lacks an explicit inductive bias for locality, leading to redundant modeling of sequence-internal local information; (ii) mixture-of-experts (MoE) implicitly couples knowledge storage with computational pathways, hindering flexible access to sequence-external global knowledge. To overcome these limitations, we propose LoKiFormer, a novel LLM architecture that augments the standard decoder with two dedicated modules: 1) Local Fusion Attention (LFA), which incorporates a convolutional fusion to attention, explicitly capturing local patterns and allowing the attention to operate on more informative representations; 2) Knowledge Memory Module (KMM), which introduces a parametric key–value memory that explicitly stores global knowledge in addressable slots, decoupling storage from computation and enabling direct knowledge retrieval. Together, these modules enable LoKiFormer to achieve more efficient and effective integration of information at both levels. Experimental results show that LoKiFormer converges 1.33x faster in pre-training than baseline models, underscoring its superiority over existing LLM architectures.