ICML 2026 Papers

Self-play bootstraps LLM reasoning through an iterative Challenger--Solver loop: the Challenger is trained to generate questions that target the Solver's capabilities, and the Solver is optimized on the generated data to expand its reasoning skills. However, existing frameworks like R-Zero often exhibit non-sustained improvement, where early gains degrade as self-play continues. We identify a key failure mode, **Diversity Illusion**, where the Solver's training signals appear diverse yet collapse into recurring underlying patterns. It manifests as (1) *Local Diversity Illusion*, where diversity is enforced only within-batch, inducing cross-iteration mode cycling; and (2) *Surface Diversity Illusion*, where questions vary superficially but require near-identical reasoning skills. To mitigate them, we propose R-Diverse with two aligned innovations: **Memory-Augmented Penalty (MAP)**, which uses a persistent memory bank to discourage recycling across iterations, and **Skill-Aware Measurement (SAM)**, which evaluates diversity by the reasoning skills exercised rather than surface variation of questions. Across 10 math and general reasoning benchmarks, R-Diverse sustains gains over more iterations and consistently outperforms prior self-play methods.

Mixture-of-Experts (MoE) models rely on balanced expert utilization to fully realize their scalability. However, existing load-balancing methods are largely heuristic and operate on noisy mini-batch assignment statistics, introducing bias relative to population-level objectives. We propose **$\phi$-balancing**, a principled framework that directly targets population-level expert balance by minimizing a strictly convex, symmetric, and differentiable potential of the expected routing distribution. Using convex duality, we derive an equivalent min-max formulation and obtain a simple online algorithm via mirror descent, yielding an efficient EMA-based routing adjustment with negligible overhead. Across large-scale pretraining and downstream fine-tuning, $\phi$-balancing consistently outperforms prior Switch-style and loss-free baselines, demonstrating more stable and effective expert utilization.

Deploying multiple models within shared GPU clusters is a key strategy to improve resource efficiency in large language model (LLM) serving. Existing multi-LLM serving systems improve GPU utilization at the cost of degraded inference performance, particularly time-to-first-token (TTFT). We attribute this degradation to the lack of awareness regarding future workload characteristics. In contrast, recent analyses have shown the strong periodicity and long-term predictability of real-world LLM serving workloads. In this paper, we propose \textit{one-for-many GPU prewarming}, which proactively loads parameters from multiple models onto GPUs based on workload forecasts. These prewarmed weights enable the system to promptly instantiate serving instances upon encountering request bursts. We design and implement WarmServe, a multi-LLM serving system incorporating three key techniques: (1) a model placement algorithm that optimizes prewarming decisions to minimize cross-model prewarming interference, (2) a KV cache reservation strategy that repurposes idle KV cache space on running GPUs for prewarming new models, and (3) an efficient GPU memory switching mechanism for tensor management. Evaluation on real-world datasets shows that WarmServe reduces tail TTFT by up to 50.8$\times$ compared to the state-of-the-art autoscaling-based system, while supporting up to 2.5$\times$ higher request throughput than the GPU-sharing system.

Vision-Language-Action (VLA) models have shown promise in robot manipulation but often struggle to generalize to new instructions or complex multi-task scenarios. We identify a critical pathology in current training paradigms where goal-driven data collection creates a dataset bias. In such datasets, language instructions are highly predictable from visual observations alone, causing the conditional mutual information between instructions and actions to vanish, a phenomenon we term Information Collapse. Consequently, models degenerate into vision-only policies that ignore language constraints. To address this, we propose LangForce, enforces instruction following via Bayesian decomposition. By introducing learnable Latent Action Queries, we construct a dual-branch architecture to estimate both a vision-only prior $p(a \mid v)$ and a language-conditioned posterior $\pi(a \mid v, \ell)$. We then optimize the policy to maximize the conditional Pointwise Mutual Information (PMI) between actions and instructions. This objective effectively penalizes the vision shortcut and rewards actions that explicitly explain the language command. Extensive experiments across on three benchmarks demonstrate substantial gains, including an 11.3\% improvement on the challenging OOD SimplerEnv benchmark, validating the ability of LangForce to robustly ground language in action.

We introduce Reinforcement Learning (RL) with Adaptive Verifiable Environments (RLVE), an approach using verifiable environments that procedurally generate problems and provide algorithmically verifiable rewards, to scale up RL for language models (LMs). RLVE enables each verifiable environment to dynamically adapt its problem difficulty distribution to the policy model's capabilities as training progresses. In contrast, static data distributions often lead to vanishing learning signals when problems are either too easy or too hard for the policy. To implement RLVE, we create RLVE-Gym, a large-scale suite of 400 verifiable environments carefully developed through manual environment engineering. Using RLVE-Gym, we show that environment scaling, i.e., expanding the collection of training environments, consistently improves generalizable reasoning capabilities. RLVE with joint training across all 400 environments in RLVE-Gym yields a 3.37% absolute average improvement across six reasoning benchmarks, starting from one of the strongest 1.5B reasoning LMs. By comparison, continuing this LM's original RL training yields only a 0.49% average absolute gain despite using over 3x more compute. We release our code publicly.

Large language model (LLM) exists a subset of attention heads that are highly responsible for long-context processing. Existing work has identified different long-context heads in models, but their detection methods mainly rely on model inference on actual long texts and do not analyze the inherent properties of the head parameters. In this paper, we use kernel methods to analyze static \textit{frequency kernels} formed by different rotation frequency components of attention heads, and we design a Long-context Potential Score (LPS) to measure the potential of attention heads in processing long contexts. Kernels of heads with high LPS exhibit concentrated low-frequency energy and low effective rank, which allow them to effectively capture highly specialized information from distant contexts. Experiments and analysis on long-context tasks and model behaviors show that the LPS metrics can well reflect the actual capability of heads on long contexts. Furthermore, by simply amplifying low-frequency kernels of heads with high retrieval potential, we can further improve model's performance on long-context tasks. Our metrics and head enhancement methods are fully static and offline, and they can be quickly conducted under low-resource constraints. Code is publicly available at [here](https://github.com/OpenDFM/long-context-heads).

The growing scale of deep neural networks, encompassing large language models (LLMs) and vision transformers (ViTs), has made training from scratch prohibitively expensive and deployment increasingly costly. These models are often used as computational monoliths with fixed cost, hindering adaptive deployment across different cost budgets. We argue that nested components, ordered by importance, can be extracted from pretrained models and selectively activated within the available computational budget. To this end, our proposed FlexRank method leverages low-rank weight decomposition with nested, importance-based consolidation to extract submodels of increasing capabilities. Our approach enables a _``train-once, deploy-everywhere''_ paradigm offering a graceful trade-off between cost and performance without training from scratch for each budget - advancing practical deployment of large models.

Fine-tuning Large Language Models (LLMs) enables data holders to construct proprietary, task-specific models by leveraging external high-performance computing infrastructure. However, existing paradigms typically address data privacy and model intellectual property (IP) in isolation, failing to simultaneously uphold both constraints. Privacy-prioritized methods compromise model IP by hosting parameters remotely, while IP-oriented collaborative schemes relying on end-to-end gradient flows inherently violate strict data privacy standards. To address these challenges, we present **PISA** (**P**rivacy-preserving and **I**P-protected **S**plit **A**daptation), a split fine-tuning framework designed to preserve both data privacy and model IP while maintaining high utility. In PISA, we propose three methods: a Manifold Rectification Pre-training (MRP) method to equip the server-side model with intrinsic robustness against privacy-induced distribution shifts; a Dual-Stream Semantic Compensation (DSC) method to recover feature utility using local clean data as priors; and a Utility-Aware Gradient Rectification (UGR) method to adaptively maximize the performance of the parameter-constrained local model. Experiments on GLUE show that PISA ensures dual protection and delivers a substantial 23.0\% performance gain over the privacy-prioritized baseline under strict privacy budgets.

A data mixture refers to how different data sources are combined to train large language models, and selecting an effective mixture is crucial for optimal downstream performance. Existing methods either conduct costly searches directly on the target model or rely on mixture scaling laws that fail to extrapolate well to large model sizes. We address these limitations by introducing a compute-efficient pipeline for data mixture scaling. First, we propose CAMEL, a capacity-aware mixture law that models validation loss with the nonlinear interplay between model size and mixture. We also introduce a loss-to-benchmark prediction law that estimates benchmark accuracy from validation loss, enabling end-to-end performance prediction for the target model. Next, we study how to allocate a fixed compute budget across model scales to fit the law and reduce prediction error. Finally, we apply our method to Mixture-of-Experts models with up to 7B-A150M parameters to fit the law, and verify the optimal mixture derived from the law by extrapolating to a 55B-A1.2B target model. Compared to prior methods, we reduce mixture optimization costs by 50\% and improves downstream benchmark performance by up to 3\%.

Despite the success of parameter-efficient fine-tuning (PEFT) methods in reducing parameter-related overhead, fine-tuning large language models (LLMs) is still bottlenecked by significant memory and computational demands. In this paper, we propose **TokenDrop**, a token-level importance-aware backpropagation skipping method that reduces activation memory and accelerates LLM fine-tuning by skipping backward computations for less informative tokens. TokenDrop evaluates token importance based on the magnitude of residual updates during the forward pass, enabling lightweight, gradient-free importance estimation. Furthermore, we introduce cumulative token selection to preserve gradient continuity across layers and lazy selection scheduling that defers token selection to facilitate globally informed importance scoring under memory constraints. Across a range of experiments, TokenDrop achieves up to **42.9**\% reduction in memory usage and up to **1.50**$\times$ training speedup, while preserving accuracy and outperforming existing backpropagation-skipping baselines. The code is available at https://github.com/kimbss470/tokendrop_official.

Mixture-of-Experts (MoE) have shown strong potential in scaling language models efficiently by activating only a small subset of experts per input. However, their deployment remains limited due to the high memory overhead associated with storing all expert parameters, particularly as the number of experts increases. To address this challenge, prior works have explored expert dropping and merging strategies; however, they often suffer from notable performance drop especially at high compression ratios due to their reliance on coarse-grained tensor- or expert-level operations. In this paper, we introduce PuzzleMoE, the first MoE merging method to enable fine-grained element-wise merging while achieving both high accuracy and inference speed, via two key innovations: First, PuzzleMoE performs sparse expert merging by identifying element-wise weight redundancy and specialization. It introduces a dual-mask approach to capture both shared and expert-specific salient parameters. Second, to avoid the overhead of storing masks and signs, we introduce a bit-packed encoding scheme that reuses underutilized exponent bits, enabling efficient MoE inference on GPUs. Extensive experiments demonstrate that PuzzleMoE outperforms prior MoE compression methods by up to 16.7\% on MMLU at 50\% compression ratio, and achieves up to 1.80$\times$ end-to-end inference throughput gain.

Tool calling extends large language models (LLMs) by enabling grounded interaction with external executable interfaces, thereby supporting environment-coupled problem solving. However, mainstream in-context learning (ICL) approaches typically incorporate detailed tool documentation and usage examples directly into the context. This results in substantial inference overhead and heightened risks of hallucination as the context length grows. Conversely, while tuning-based methods improve general tool-calling capabilities, they often fail to effectively internalize the specific details of previously seen tools, thereby retaining a dependency on in-context documentation. To address these limitations, we propose ParaTool, a framework that projects each tool into a dedicated, loadable set of parameters. By equipping a dynamic integration of these parameterized tools, the LLM can perform tool calling without relying on in-context documents or examples. Specifically, our approach consists of three stages: (1) parametric tool pre-training encapsulates the knowledge of different tools into independent parameter modules; (2) soft tool selection employs a gating network to dynamically weigh and aggregate relevant tool parameters; and (3) parametric tool fine-tuning jointly updates tool parameters to align the training and inference processes. Experiments on Stable ToolBench and BFCL demonstrate that ParaTool significantly outperforms strong ICL-based baselines, achieving superior performance while reducing computational complexity.

Sparse Mixture-of-Experts (MoE) language models enable conditional computation but face deployment challenges due to the memory wall: while few experts are activated per token, the entire model must reside in memory. Existing expert pruning methods primarily rely on independent ranking, failing to account for the complex inter-dependencies and redundancies between experts. In this paper, we formulate post-training MoE pruning as a reconstruction-driven subset selection problem, aiming to minimize layer-output distortion under a cardinality constraint. We introduce SCHUR-A*, an algorithm that leverages A* search to achieve globally optimal expert selection within each layer. To maintain computational tractability, we derive a novel, admissible heuristic upper bound using a Schur-complement-based relaxation of the reconstruction objective. This tight bound allows for aggressive pruning of the search space while mathematically guaranteeing optimality. Furthermore, we propose an automated strategy to balance fidelity and memory reduction across heterogeneous layers via knee-point detection. Extensive experiments on Qwen3-30B-A3B demonstrate that SCHUR-A* significantly outperforms greedy and ranking-based baselines, maintaining comparable performance even under aggressive pruning ratios.

Foundation model-driven agents often struggle with long-horizon planning due to the transient nature of purely prompting-based reasoning. While existing skill induction methods mitigate this by distilling experience into state-blind parameterized scripts, they fail to capture the conditional logic required for robust execution in dynamic environments. In this paper, we propose Neuro-Symbolic Skill Induction (NSI), a framework that lifts interaction traces into modular, logic-grounded programs. By synthesizing explicit control flows and dynamic variable binding, NSI empowers agents to discover when and why to act. This paradigm enables the efficient generalization, allowing agents to induce skills from few-shot examples and flexibly adapt to unseen goals. Experiments on a series of agentic tasks demonstrate that NSI consistently outperforms state-of-the-art baselines, empowering agents to self-evolve into architects of logic-grounded skills. Project Page: https://sh-jj.github.io/NSI.

While existing multi-agent systems (MAS) can handle complex problems by enabling collaboration among multiple agents, they are often highly task-specific, relying on manually crafted agent roles and interaction prompts, which leads to increased architectural complexity and limited reusability across tasks. Moreover, most MAS communicate primarily through natural language, making them vulnerable to error accumulation and instability in long-context, multi-stage interactions within internal agent histories. In this work, we propose \textbf{Agent Primitives}, a set of reusable latent building blocks for LLM-based MAS. Inspired by neural network design, where complex models are built from reusable components, we observe that many existing MAS architectures can be decomposed into a small number of recurring internal computation patterns. Based on this observation, we instantiate three primitives (Review, Voting and Selection, and Planning and Execution), all communicating via key–value (KV) cache to mitigate information degradation across multi-stage interactions. To enable automatic system construction, an Organizer agent automatically selects and composes primitives for each query, guided by a lightweight knowledge pool of previously successful configurations, forming a primitive-based MAS. Experiments show that primitives-based MAS improve average accuracy by 12.0–16.5\% over single-agent baselines, reduce token usage and inference latency by approximately 3$\times$–4$\times$ compared to text-based MAS, while incurring only 1.3$\times$–1.6$\times$ overhead relative to single-agent inference and providing more stable performance across model backbones.

A reliable reward model is essential for aligning large language models (LLMs) with human preferences through reinforcement learning from human feedback (RLHF). However, standard reward models are susceptible to spurious features that are not causally related to human labels. This can lead to *reward hacking*, where high predicted reward does not translate into better behavior. In this work, we address this problem from a causal perspective by proposing a factored representation learning framework that decomposes the model’s contextual embedding into (1) causal factors that are sufficient for reward prediction and (2) non-causal factors that capture reward-irrelevant attributes such as length or sycophantic bias. The reward head is then constrained to depend only on the causal component. In addition, we introduce an adversarial head trained to predict reward from the non-causal factors, while applying gradient reversal to discourage them from encoding reward-relevant information. Experiments on both mathematical and dialogue tasks demonstrate that our method learns more robust reward models and consistently improves downstream RLHF performance over state-of-the-art baselines. Analyses on length and sycophantic bias further validate the effectiveness of our method in mitigating reward hacking behaviors.

Mode collapse is a persistent challenge in generative modeling and manifests in autoregressive text generation as behaviors ranging from explicit looping to gradual loss of diversity and premature trajectory convergence. We take a dynamical-systems view and reinterpret mode collapse as reduced state-space accessibility caused by *geometric collapse*: during generation, the model's internal trajectory becomes confined to a low-dimensional region of its representation space. This implies mode collapse is not purely a token-level phenomenon and cannot be reliably mitigated by symbolic constraints or probability-only decoding heuristics. Guided by this perspective, we propose *Reinforced Mode Regulation* (RMR), a lightweight, online state-space intervention that regulates dominant self-reinforcing directions in the Transformer value cache (implemented as low-rank damping). Across multiple large language models, RMR substantially reduces mode collapse and enables stable, high-quality generation at extremely low entropy rates (down to 0.8 nats/step), whereas standard decoding typically collapses near 2.0 nats/step.

Hallucinations pose a key challenge for large language models, and chain-of-thought prompting exposes intermediate reasoning but usually treats traces as linear sequences, making crossstep dependencies and unsupported intermediate claims difficult to identify. We propose a structural reasoning model to describe interactions among local reasoning steps. To detect hallucinations, we extract a directed acyclic reasoning graph over conditions and intermediate claims, verify each claim against its parent nodes, and aggregate the step signals with a simple mass-flow rule. Under a probabilistic erasure-gate abstraction, we interpret this aggregation as measuring information loss along the reasoning graph. Experiments on GSM8K, MATH, HumanEval, and HotpotQA show that the proposed method is most advantageous on longer, dependency-rich reasoning traces such as math and code generation, while remaining competitive on shorter factual QA; these results provide a structured perspective on chain-of-thought evaluation. All code and data are available at https://github.com/soncheinbok/FlowScore.

Reinforcement Learning from Human Feedback (RLHF) has become the standard for aligning Large Language Models (LLMs), yet its efficacy is bottlenecked by the high cost of acquiring preference data, especially in low-resource and expert domains. To address this, we introduce ActiveUltraFeedback, a modular active learning pipeline that leverages uncertainty estimates to dynamically identify the most informative responses for annotation. Our pipeline facilitates the systematic evaluation of standard response selection methods alongside Double Reverse Thompson Sampling (DRTS) and DeltaUCB, two novel methods prioritizing response pairs with large predicted quality gaps, leveraging recent results showing that such pairs provide good signals for fine-tuning. Our experiments demonstrate that ActiveUltraFeedback yields high-quality datasets that lead to significant improvements in downstream performance, notably achieving comparable or superior results with as little as one-sixth of the annotated data relative to static baselines. Our pipeline is available at https://github.com/lasgroup/ActiveUltraFeedback and our preference datasets at https://huggingface.co/ActiveUltraFeedback.