The paper proposes a method for selectively querying language‑model advice in reinforcement learning by predicting the value of potential responses and only querying when the expected benefit outweighs the cost. It introduces a certified, response‑contingent metareasoning framework that guarantees near‑optimal advice usage under certain assumptions, and demonstrates that a calibrated controller with Qwen2.5 advisors can improve task performance while drastically reducing the number of advice calls on the BabyAI benchmark.
By Ibne Farabi Shihab, Md Najmus Swaqeeb, Abu Sa-Adat Mohamed Moon-Im Al Ahsan
arXiv:2608. 15565v1 Announce Type: new Abstract: Experience-learning agents for optimization modeling improve by storing verified skills, but existing learners admit knowledge by checking against known answers, which real ticket streams do not provide.
By Junbo Jacob Lian, Huiling Chen, Hanzhang Qin, Chung-Piaw Teo
SkillLift introduces a method for efficiently evolving reusable procedural prompts (skills) in large language model agents by learning a dense rubric that aligns with sparse oracle evaluations. Instead of directly revising skill text based on costly full agent rollouts, the approach decouples skill search from oracle cost through a bilevel optimization framework: an inner loop uses a frozen rubric as a cheap surrogate to guide skill updates, while an outer loop periodically realigns the rubric using a small number of oracle rollouts via rank correlation. Experiments on complex agent task benchmarks demonstrate that SkillLift outperforms existing auto-skill methods while reducing token cost by 40–70% compared to frontier-evolving approaches.
By Haoxiang Kang, Ming Wen
The paper investigates how trajectory fine‑tuning can enhance small language models (SLMs) as next‑action controllers in retrieval‑augmented question answering. By building a seven‑way action‑prediction task from teacher search traces, the authors fine‑tune SLMs and cross‑lingual SLMs (xSLMs) using LoRA and evaluate on 1,646 held‑out examples, achieving a macro‑F1 of 0.6536 with Granite 4.1 3B. In an end‑to‑end controller/generator swap experiment on 149 trajectories, the fine‑tuned model improves Exact Match from 0.7530 to 0.7946 and token F1 from 0.7783 to 0.8295, demonstrating that trajectory supervision boosts action prediction and evidence‑recording behavior.
By Mohammed Al-Maamari, Saber Zerhoudi, Michael Granitzer, Jelena Mitrovi\'c
MInTRL (Minimal Intervention Reinforcement Learning) expands exploration in on-policy reinforcement learning by inserting sparse, local corrections into rollouts via a judge-intervention policy. These interventions replace erroneous suffixes and immediately return control to the main policy, allowing the agent to explore beyond its natural trajectory while maintaining on-policy data. The method uses a sequence-level advantage-regression objective, avoiding importance sampling, and demonstrates superior performance on math and code benchmarks compared to standard on-policy and off-policy baselines.
By Mingyu Chen, Yefan Tao, Gerald Friedland, Xuezhou Zhang, Chris Kong
arXiv:2606. 04161v1 Announce Type: new Abstract: Different predictors often excel on different inputs, so picking the best one per instance promises higher accuracy than committing to a single model.
By Tyler Crosse, Alan Nadelsticher Ruvalcaba, Dustin Khang LeDuc, Thomas Trask, Nicholas Lytle, David Joyner
The paper investigates two strategies for improving large language model (LLM) evaluation: specialized judge weights and rule‑based deferral policies. Experiments on nearly 100,000 rubric‑conditioned samples show that correct rubrics boost accuracy, while incorrect ones hurt it, and that splitting training data into criterion‑specific experts can severely degrade performance unless the experts are warm‑started from a unified model. The authors demonstrate that lightweight deferral cascades can match or exceed the accuracy of larger standalone judges at a fraction of the compute cost, and they provide practical design rules for building efficient, reliable LLM evaluators.
By Ye Chen, Weining Zhang
GEAR is a two‑stage framework that distills tabular foundation models into lightweight MLP or tree‑based predictors for efficient CPU deployment. In the first stage, synthetic covariates are used as teacher‑query locations to train the student on soft TFM targets, expanding coverage beyond observed rows. The second stage re‑anchors the student to the target distribution using real labels and out‑of‑fold teacher predictions, preventing self‑labeling leakage and improving performance. Experiments on TALENT and TabArena show that GEAR‑distilled MLPs outperform supervised MLPs by up to 2.00 AUC points on binary tasks and 1.35 on multiclass tasks, and also outperform CatBoost, while dramatically reducing inference time and memory usage.
By Qi Qin, Jiajie Zhu, Dali Chen, Yuzhao Zhang, Jia-Xing Han, Yu Su, Peng Zhang, Ying Yan, Yifan Sun
Instella‑MoE is a fully open Mixture‑of‑Experts language model with 16 billion total parameters and 2.8 billion active parameters per token, trained from scratch on AMD Instinct GPUs. It incorporates a sparsely activated MoE design with Gated Multi‑head Latent Attention and FarSkip‑Collective connectivity, and follows a multi‑stage pipeline that includes pre‑training, long‑context extension, supervised fine‑tuning, direct preference optimization, and reinforcement learning with Multi‑Teacher On‑Policy Distillation. The model achieves an average score of 76.7 on pre‑training benchmarks and 73.2 on instruction‑following, reasoning, math, coding, and chat benchmarks, outperforming comparable fully open and open‑weight models, and its full training pipeline, weights, and code are released for reproducibility.
By Jiang Liu, Sudhanshu Ranjan, Prakamya Mishra, Yonatan Dukler, Gowtham Ramesh, Jialian Wu, Ximeng Sun, Wen Xie, Chaojun Hou, Vikram Appia, Zhenyu Gu, Zicheng Liu, Emad Barsoum
The paper investigates on‑policy distillation (OPD) of large language models when trained on a minimal amount of data, specifically a single query. It finds that one‑shot OPD continues to improve for hundreds of steps and recovers most of the performance gains achieved with full‑dataset training across various tasks and model families. The authors attribute this to the high state coverage achieved early in training and show that adding a few semantically distinct queries further boosts coverage and accuracy, eventually matching full‑data results. They conclude that OPD is data‑overfed but algorithm‑starved, highlighting the need to focus on step efficiency and the underlying mechanisms of its recent successes.
By Zixuan Fu, Bingxiang He, Yuxin Zuo, Haohuan Huang, Jinqian Zhang, Ruhang Xiao, Cheng Qian, Qinyu Luo, Huan-ang Gao, Yudong Wang, Zhiyuan Liu, Ning Ding, Chaojun Xiao
Large language models (LLMs) are increasingly deployed in complex, compound AI systems where performance hinges on the quality of prompts. Recent state-of-the-art optimizers like GEPA (Genetic-Pareto) have argued that reflective instruction evolution can outperform traditional reinforcement learning and few-shot optimization.
arXiv:2606. 09078v1 Announce Type: new Abstract: Process Reward Models (PRMs) improve credit assignment for reasoning by providing step-level feedback.
By Aakriti Agrawal, Souradip Chakraborty, Armin Saghafian, Nihal Sharma, Rizal Fathony, Nam H Nguyen, C. Bayan Bruss, Amrit Singh Bedi, Furong Huang