The paper shows that fine‑tuning a large language model (LLM) with a debate framework—where a generator and a critic compete and a weaker LLM judge adjudicates—reduces reward hacking compared to standard reinforcement learning from AI feedback (RLAIF). In experiments on mathematics tasks, the debate approach keeps the judge’s performance stable, achieving a 45% higher peak validation accuracy than the RLAIF baseline and mitigating the rapid exploitation of judge errors. Additional findings indicate that weakening the judge speeds hacking unless countered by extra debate rounds, that debate can override misalignment prompts, and that word‑limit constraints on critiques help balance the game and prevent judge hacking.
whyItMatters":"The study demonstrates a practical method to curb reward hacking in RL‑based AI systems, addressing a key obstacle for safely scaling AI oversight."
By Zachary Kenton, Lili Janzer, Rory Greig, Tian Huey Teh, Kirill Tyshchuk, Jonah Brown-Cohen, Harri Edwards, Senthooran Rajamanoharan, Noah Y. Siegel, Natasha Jaques, Rohin Shah
arXiv:2607. 25308v1 Announce Type: cross Abstract: Training large language models (LLMs) to act in long-horizon games is a promising step toward generalist decision-making, yet reinforcement learning with verifiable rewards (RLVR) relies on sparse final rewards that reveal little about which decisions determine success.
By Yu Wang, Yi-Kai Zhang, Wentao Shi, Ziang Ye, Yuchun Miao, Yueqing Sun, Qi Gu, Xunliang Cai, Lan-Zhe Guo, Han-Jia Ye, Fuli Feng
The paper explores a runtime strategy-selection framework where a large language model (LLM) guides a pre‑trained reinforcement learning (RL) policy for non‑player characters (NPCs) in a Unity combat game without altering the underlying policy. Five NPC agents sharing a PPO policy were compared in a baseline setup and an LLM‑augmented setup, where a locally hosted Mistral 7B model assigns one of four tactical tags every five seconds based on live game state. Across 600 episodes against three scripted opponents, the LLM‑augmented agents more than doubled their win rate against a Balanced opponent, improved performance against an Evasive opponent, but struggled against an Aggressive opponent due to over‑reliance on encirclement; analysis of 2,430 strategy selections revealed limited zero‑shot differentiation with the model favoring Surround in 83.8% of cases.
By Hrithika Deepu Nair, Kayvan Karim
arXiv:2607. 05904v1 Announce Type: new Abstract: Training a language model against its own reference-free judgments (the premise of self-rewarding, self-play, and LLM-as-a-judge pipelines) assumes a model's verdict on a shown answer tracks correctness.
By Chenyu Zhou
The paper introduces TalkMesh, a decentralized network of small language model agents that learn to communicate effectively during inference. Each agent proposes an answer, scores it with a confidence head, and the most confident agent broadcasts a hint; lower‑confidence agents revise their proposals if a new suggestion scores higher. This gossip‑based consensus, trained via group relative policy optimization, enables a mesh of three agents to match the accuracy of majority voting over 32 samples, and scales to larger meshes to significantly boost performance on benchmarks like GSM8K and MATH-500.
By Mehmet Kerem Turkcan
arXiv:2607. 18006v1 Announce Type: cross Abstract: Large language models achieve strong reasoning performance, but often at prohibitive training cost - a challenge that is especially acute for compact models ($\leq 4 \, \mathrm{B}$ parameters) trained under limited budgets.
By Martino M. L. Pulici, Cuong Xuan Chu, Evgeny Kharlamov, Zifeng Ding, Volker Tresp, Yunpu Ma
arXiv:2607. 00190v1 Announce Type: cross Abstract: Recent advances in reinforcement learning have produced superhuman agents across a wide range of competitive games.
By Andrzej Bia{\l}ecki, Adam Mastalerz, Han Zhou
Large language models achieve strong reasoning performance, but often at prohibitive training cost - a challenge that is especially acute for compact models ($\leq 4 \, \mathrm{B}$ parameters) trained under limited budgets. We introduce MADA-RL, a post-training framework that specializes compact models into generator and critic roles and trains them with a debate-aware learning signal, fine-tuning only a small subset of parameters via LoRA adapters.
arXiv:2607. 21856v1 Announce Type: new Abstract: Modern reasoning models depend on reasoning data, today sourced from human annotations or distilled from stronger LLMs.
By Ziran Yang, Chengshuai Shi, Raj Ghugare, Benjamin Eysenbach, Karthik Narasimhan, Chi Jin
SkillGate is a method that trains agents to select the correct skill from a large slate during an episode by separating credit signals for skill selection and execution. It addresses the problem of selector credit starvation, where traditional outcome-rewarded RL fails to give sufficient credit to the skill-naming tokens, especially in long-horizon tasks. Experiments on five benchmarks show that SkillGate improves a 9B policy’s success rate from 40.8% to 53.2%, reduces exposure to misleading candidates, and requires fewer skill reads.
By Qingyao Li, Wenxiang Jiao, Shuai Shao, Kangning Zhang, Yuan Lu, Yi Guo, Weiwen Liu, Weinan Zhang, Yong Yu
The paper introduces the Agentic Compositional Generalization hypothesis, suggesting that reinforcement learning (RL) primarily refines high‑level decision‑making behaviors that orchestrate pre‑trained low‑level skills, rather than teaching new domain‑specific skills from scratch. It proposes River, a training recipe that enhances reward quality by filtering low‑quality synthetic environments and adding process‑level behavior regularization. Using River, RL‑trained agents outperform other open‑source 8B models on four terminal‑agent benchmarks, achieving significant gains with fewer than 30% of the training environments.
By Yihang Yao, Bo Pang, Xuan Phi Nguyen, Ding Zhao, Shafiq Joty, Semih Yavuz
The paper introduces Local Updates, Global Learning (LUGL), a framework that separates data collection from model training, allowing non‑incremental learners such as gradient‑boosted trees (LightGBM) to be used in reinforcement learning for games. LUGL alternates between a local phase—where self‑play generates tabular updates—and a global phase—where these updates train a function approximator before resetting the table. Experiments on both perfect‑information and imperfect‑information games show that LightGBM agents perform competitively or better than neural‑network baselines like DQN and DeepCFR.