1
Diffract: Spectral View of LLM Domain Adaptation
用谱视角剖析大模型领域适应，Diffract方法为持续预训练提供新洞察。
arXiv:2608.10850v1 Announce Type: new Abstract: We study continual pre-training (CPT) as a mechanism for adapting general-purpose large language model…
2
Stem: Rethinking Causal Information Flow in Sparse Attention
稀疏注意力中因果信息流被重新定义，Stem提出新视角，为长上下文模型效率与准确性兼得提供关键思路。
arXiv:2603.06274v2 Announce Type: replace-cross Abstract: The quadratic computational complexity of self-attention remains a fundamental bottleneck fo…
3
Multi-Task GRPO: Reliable LLM Reasoning Across Tasks
提出多任务GRPO方法，有效提升大模型跨任务推理的可靠性，已被ICML 2026收录。
arXiv:2602.05547v2 Announce Type: replace-cross Abstract: RL-based post-training with GRPO is widely used to improve large language models on individu…
4
Leveraging Biokinetic Knowledge Priors for Data-Scarce Bioprocess Modeling
用生物动力学知识先验解决数据稀缺难题，为生物过程建模开辟新路径——ICML 2026 AI for Science 论文
arXiv:2607.20539v1 Announce Type: cross Abstract: While deep learning has accelerated drug discovery, its impact on biomanufacturing has been consider…
5
Concept Concentration for Faithful Representation Intervention
ICML'26提出概念集中方法，通过忠实表示干预提升AI可解释性，为模型控制提供新思路。
arXiv:2505.18672v2 Announce Type: replace Abstract: Representation intervention aims to localize and modify the representations that encode the underl…
6
Reward-Aware Population Scaling of Evolutionary Strategies in LLM Fine-Tuning
新方法通过奖励感知群体缩放优化进化策略，显著提升LLM微调效率，已被ICML 2026 workshop收录。
arXiv:2607.19408v1 Announce Type: new Abstract: Using Evolutionary Strategies (ES) for fine-tuning large language models is attractive because it is m…
7
Efficient Clustering with Provable Guardrails for LLM Inference at Scale
提出可证明安全的高效聚类方案，为大规模LLM推理保驾护航。
arXiv:2607.19704v1 Announce Type: new Abstract: Scaling LLM-based applications to millions of users is bottlenecked by the inference cost and latency …
8
From Outcomes to Actions: Leveraging Hindsight for Long-Horizon Language Agent Training
提出利用事后经验将结果转化为行动，显著提升长周期语言智能体训练效果，ICML 2026录用论文。
arXiv:2607.16257v1 Announce Type: cross Abstract: Reinforcement learning (RL) has become a widely adopted technique for improving large language model…
9
Stabilizing Native Low-Rank LLM Pretraining
低秩预训练大模型遭遇不稳定困境？这项ICML 2026研究提出原生低秩LLM预训练的稳定化方法，兼顾效率与质量。
arXiv:2602.12429v2 Announce Type: replace Abstract: Foundation models have achieved remarkable success, yet their growing parameter counts pose signif…
10
Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning
ICML 2026重磅论文：进化策略替代强化学习，开创大模型微调新范式。
arXiv:2509.24372v3 Announce Type: replace Abstract: Fine-tuning large language models (LLMs) for downstream tasks is an essential stage of modern AI d…
11
Are we Merging the Right Models? Impact of Expert Training Duration on Model Merging for LLMs
专家训练时长如何影响LLM模型合并效果？这篇ICML 2026 workshop论文揭示关键发现。
arXiv:2607.11997v1 Announce Type: new Abstract: Multi-task model merging combines separately trained expert models into a single model that handles al…
12
Before the Last Token: Diagnosing Final-Token Safety Probe Failures
揭秘大模型安全探针为何在最终token失效：早于最后一层的内部表征已暴露漏洞
arXiv:2605.12726v2 Announce Type: replace Abstract: Final-token safety probes monitor a single hidden state after prompt prefill, but jailbreak prompt…
13
Towards Understanding Steering Strength
ICML 2026论文深入解析模型引导技术中的强度量化问题，为AI可解释性提供新视角。
arXiv:2602.02712v2 Announce Type: replace Abstract: A popular approach to post-training control of large language models (LLMs) is the steering of int…
14
Dissociating the Internal Representations of Sycophancy in LLMs
从神经层面拆解大模型「拍马屁」行为的内部机制，一篇被ICML 2026研讨会接收的机械可解释性突破。
arXiv:2607.07003v1 Announce Type: new Abstract: Large Language Models (LLMs) frequently exhibit sycophancy, where they agree with a user's statement e…
15
RL Post-Training Builds Compositional Reasoning Strategies
强化学习后训练如何催生组合推理策略，来自ICML workshop的前沿实证。
arXiv:2607.07646v1 Announce Type: new Abstract: Does RL post-training merely amplify primitive skills already latent in a base model, or can it compos…
16
Beyond Heuristic Tuning: Power-Calibrated LLM Watermarking
ICML 2026接收：为LLM水印引入功率校准，告别传统启发式调优。
arXiv:2607.05694v1 Announce Type: cross Abstract: Logit-based watermarking is a widely used mechanism for identifying LLM generated content, yet its e…
17
CoopEval: Benchmarking Cooperation-Sustaining Mechanisms and LLM Agents in Social Dilemmas
ICML 2026论文为LLM Agent在社交困境中的合作能力打造首个系统化基准，揭示维持机制的关键设计。
arXiv:2604.15267v2 Announce Type: replace-cross Abstract: It is increasingly important that LLM agents interact effectively and safely with other goal…
18
Language Generation with Replay: A Learning-Theoretic View of Model Collapse
从学习理论视角解析模型崩溃，重放机制能否成为破局关键？ICML 2026 新作给出严谨答案。
arXiv:2603.11784v2 Announce Type: replace Abstract: As scaling laws push the training of frontier large language models (LLMs) toward ever-growing dat…
19
The Moving Target: A Longitudinal Audit of Trustworthiness Drift Across Twelve Checkpoints of Open-Source Chat LLMs
追踪开源聊天LLM十二个版本的信任度变化，揭示模型行为随时间漂移的潜在风险
arXiv:2607.02587v1 Announce Type: cross Abstract: Model cards quote trust-benchmark scores without recording when they were measured, and the same num…
20
ARLArena: A Unified Framework for Stable Agentic Reinforcement Learning
统一框架稳定智能体强化学习，ICML 2026收录，直击训练不稳定痛点。
arXiv:2602.21534v3 Announce Type: replace Abstract: Agentic reinforcement learning (ARL) has rapidly gained attention as a promising paradigm for trai…