OrchRM: Self-supervised reward modeling for multi-agent orchestration without human annotations
Researchers propose Orchestration Reward Modeling (OrchRM), a self-supervised framework that trains reward models for LLM-based multi-agent orchestrators using intermediate execution artifacts to construct win-lose pairs for Bradley-Terry training. The approach avoids costly sub-agent rollouts by operating directly at the orchestration level, achieving up to 10x improvement in training token efficiency and up to 8% accuracy gains in test-time scaling. Results generalize across mathematical reasoning, web-based QA, and multi-hop reasoning tasks.
Related guides (2)
Related events (8)
Skill-RM: A unified reward model framework treating evaluation as an agentic skill
Researchers from the Qwen team propose Skill-RM, a framework that reformulates reward modeling as the execution of a reusable 'Reward-Evaluation Skill,' enabling a single model to orchestrate heterogeneous evaluation criteria including rule-based verifiers, ground-truth references, and rubrics. By treating reward computation as a structured agentic task, Skill-RM dynamically selects and aggregates evidence per input rather than relying on static evaluation. Experiments on reward benchmarks and downstream tasks (best-of-N selection, RL) show consistent improvements over traditional judge baselines. The code is publicly released under the Qwen-Applications GitHub organization.
ExpRL: RL-based mid-training using human QA data as reward scaffolds for LLM reasoning
ExpRL proposes an automated approach to LLM mid-training that replaces manually curated reasoning traces with large corpora of human-written QA data used as reward scaffolds rather than imitation targets. Reference solutions are hidden from the policy and used only to construct problem-specific grading rubrics, enabling dense process-level rewards that reinforce partial progress and intermediate reasoning steps. On challenging math reasoning benchmarks, ExpRL outperforms SFT, sparse-reward GRPO, and self-distillation as an RL initialization strategy, with additional mixed-domain experiments suggesting broader applicability.
Improving Model Safety Behavior with Rule-Based Rewards
OpenAI has developed a method called Rule-Based Rewards (RBRs) that trains models to behave safely without requiring extensive human data collection. The approach uses explicit rules to generate reward signals during training, offering a more scalable alternative to traditional RLHF-based safety alignment. This represents a practical contribution to alignment methodology from a Tier 1 lab.
Two is better than one: A Collapse-free Multi-Reward RLIF Training Framework
This paper proposes a multi-reward reinforcement learning from internal feedback (RLIF) framework that decomposes training signals into an answer-level reward via cluster voting and a completion-level reward via token-wise self-certainty. To address reward hacking and entropy collapse common in single-reward RLIF, the authors introduce GDPO-based normalization and KL-Cov regularization targeting low-entropy token distributions. Evaluated on mathematical reasoning and code-generation benchmarks, the method achieves stability and performance approaching supervised RLVR methods without requiring external ground-truth supervision. The work advances scalable unsupervised RL training for LLM reasoning.
Learning from Human Preferences: OpenAI and DeepMind Collaborate on Reward Learning from Comparisons
OpenAI, in collaboration with DeepMind's safety team, published a method for learning reward functions directly from human preference comparisons between pairs of agent behaviors, eliminating the need to hand-code goal functions. The algorithm infers human intent by asking evaluators which of two proposed behaviors is preferable, addressing risks from misspecified reward functions. This work is an early foundational contribution to what would become reinforcement learning from human feedback (RLHF). It targets both safety and alignment concerns around reward hacking and proxy gaming.
PROVE framework trains LLMs for multi-step tool use via stateful MCP environments and programmatic rewards
Researchers introduce PROVE (Programmatic Rewards On Verified Environments), a framework for training LLMs to orchestrate multi-step tool calls using reinforcement learning. The system includes a library of 20 stateful MCP servers with 343 tools, an automated data synthesis pipeline that grounds training queries in live server state, and a multi-component programmatic reward function requiring no judge model. Training four models (Qwen3-4B, Qwen3-8B, Qwen2.5-7B, Granite-4.1-8B) with ~13K examples yields gains of up to +10.2 on BFCL Multi-Turn, +6.8 on tau2-bench, and +6.5 on T-Eval, demonstrating consistent improvements in multi-step tool orchestration.
PCMA: Learning coordinated agent-specific preferences for multi-objective multi-agent RL
A new arXiv preprint introduces Preference Coordinated Multi-agent Policy Optimization (PCMA), a method for cooperative multi-objective multi-agent reinforcement learning (MOMARL) that learns agent-specific preferences to enable complementary trade-offs across agents. The authors formulate cooperative MOMARL as a team-optimal game and provide a first-order improvement decomposition showing that preference diversity can induce team improvement. Experiments on cooperative MOMA environments and a traffic-control scenario demonstrate improvements in both performance and trade-off coordination.
AgenticRL: Self-refining LLM-guided reward design and policy refinement for UAV navigation
AgenticRL is a framework that uses a multimodal GPT agent to automate reward function generation, policy training via PPO, and closed-loop self-refinement for UAV navigation tasks. The agent evaluates trained policies through diagnostic feedback, identifies failure modes, and iteratively refines rewards without human intervention. Evaluated across five navigation tasks, the closed-loop refinement improves policy behavior by 71% over initial rewards, with sim-to-real transfer achieving 91% real-world success rate and 94% sim-to-real accuracy.