Fine-tunes long-horizon LLM agents with evolution strategies so full-model updates run at inference-level GPU memory. Emphasizes trajectory-level credit via black-box rewards, online prompt–parameter co-evolution, and a cosine decay for perturbation scale to balance exploration and adaptation; suited for limited-GPU settings.
Estimates optimal learning rates for large-scale Mixture-of-Experts pretraining using a two-step, compute-efficient transfer: μP-based width transfer from small proxy models, then log-log linear extrapolation across token budgets to trillion-token horizons.
AI Video Papers2026
Yunheng Li, Guohong Mu +5VCIP, School of Computer Science, Nankai University, Brain and Artificial Intelligence Lab, Northwestern Polytechnical University +2
Treats human annotations as oracle rollouts and separates them from on-policy baselines to improve reinforcement learning for video multimodal LLMs. Key features include a decoupled advantage estimator, sign-balanced pruning, and scalable gains across model sizes and data budgets.
AI Agent Papers2026
Sungho Park, Wonjoong Kim +11Affiliation: KAIST, Affiliation: Southern University of Science and Technology +1
Automatically optimizes runtime harnesses for LLM agents by diagnosing failure traces and iteratively applying structured, generalizable patches. Combines batch-based failure diagnosis, code-like patch generation across prompts/tools/middleware, and validation-aware selection to raise long-horizon task success on multiple benchmarks.
Analyzes on-policy self-distillation for language-model reasoning, diagnosing “collapse” and framing it as controlled by three levers: where token-level signals apply, what privileged information the teacher sees, and how teacher dynamics evolve.
AI Agent Papers2026
Synthesizes, repairs, and self-evolves task-adaptive agent harnesses on demand for off-the-shelf LLM agents, using a trainable harness-intelligence model that distills signals from past configurations. Demonstrates consistent performance gains across benchmarks and model families by producing four-module, composable harnesses.
AI Agent Papers2026
Analyzes how to generate useful interaction data for LLM agents and proposes the ACE lens — Accuracy, Complexity, divErsity — while factorizing agentic data as (E, q, τ, v). Surveys verification, difficulty calibration, and coverage strategies and outlines implications for training and benchmarks.
A zero-data self-evolution framework that co-trains a Challenger, Solver, and Judge so LLMs can iteratively improve on both verifiable and unverifiable tasks without human labels. Uses role-asymmetry and subtask-amplification preference pairs to train the Judge and sustain improvement.
Aozhe Wang, Zhengxi Lu +9Affiliation: Zhejiang University, Affiliation: Alibaba Group{waz,zhengxilu,syl}@zju.edu.cn [email protected]
A test-time method that adapts LLMs without labels by distilling rollouts that agree with majority pseudo-labels and penalizing disagreeing rollouts via grouped RL, improving robustness under frequent pseudo-label errors.
Decides when past post-training updates should be reused for autonomous LLM adaptation by introducing Boundary-Calibrated Intervention Transfer (BCIT). BCIT binds effects to source context, checks applicability and hard conflicts, and runs bounded trials to obtain current-state evidence—reducing harmful updates and improving equal-budget final-model quality.
Analyzes on-policy distillation for LLM fine-tuning, shows teacher token-level supervision is often noisy and not the main driver of gains, and introduces OPSA, a supervision-free, entropy-adaptive method that suppresses low-probability tokens to improve downstream accuracy.
Studies looping shared transformer layers in Mixture-of-Experts models under matched budgets and proposes SMELT: loop the middle half twice while matching per-token FLOPs, non-embedding parameters, and KV cache. Shows 6.8–18.0% training-FLOPs savings on the compute-optimal frontier, stronger downstream gains on code and long-context tasks.