REVIEW 19 cited by
ArCHer: Training Language Model Agents via Hierarchical Multi-Turn RL
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
A broad use case of large language models (LLMs) is in goal-directed decision-making tasks (or "agent" tasks), where an LLM needs to not just generate completions for a given prompt, but rather make intelligent decisions over a multi-turn interaction to accomplish a task (e.g., when interacting with the web, using tools, or providing customer support). Reinforcement learning (RL) provides a general paradigm to address such agent tasks, but current RL methods for LLMs largely focus on optimizing single-turn rewards. By construction, most single-turn RL methods cannot endow LLMs with the ability to intelligently seek information over multiple turns, perform credit assignment, or reason about their past actions -- all of which are critical in agent tasks. This raises the question: how can we design effective and efficient multi-turn RL algorithms for LLMs? In this paper, we develop a framework for building multi-turn RL algorithms for fine-tuning LLMs, that preserves the flexibility of existing single-turn RL methods for LLMs (e.g., proximal policy optimization), while accommodating multiple turns, long horizons, and delayed rewards effectively. To do this, our framework adopts a hierarchical RL approach and runs two RL algorithms in parallel: a high-level off-policy value-based RL algorithm to aggregate reward over utterances, and a low-level RL algorithm that utilizes this high-level value function to train a token policy within each utterance or turn. Our hierarchical framework, Actor-Critic Framework with a Hierarchical Structure (ArCHer), can also give rise to other RL methods. Empirically, we find that ArCHer significantly improves efficiency and performance on agent tasks, attaining a sample efficiency of about 100x over existing methods, while also improving with larger model capacity (upto the 7 billion scale that we tested on).
Forward citations
Cited by 19 Pith papers
-
The Dark Room in the Reward Channel: Dense Prediction Rewards Collapse GRPO-Trained LLM Agents -- and The Channel, Not the Content, Decides What Works
A potential-based prediction reward collapses GRPO-trained LLM agents into a predictable 'dark room' state, and the collapse is caused by GRPO's std normalization rather than by the reward's magnitude.
-
Instruction-Conditioned Exploration for Reinforcement Learning with Self-Distillation to an Unconditioned Policy
Training with appended behavioral instructions plus correctness-filtered self-distillation improves held-out math pass@1 over DAPO for a 1.7B model, but not for 4B at 4K context.
-
Teaching LLMs to Self-Evolve: Cultivating Core Meta-Skills with Reinforcement Learning
MetaEvolve trains LLMs with reinforcement learning on synthesized code-refinement trajectories, reporting large gains on coding and numerical-optimization benchmarks.
-
ABot-AgentOS: A General Robotic Agent OS with Lifelong Multi-modal Memory
A robotic agent operating system with source-grounded graph memory and split-wise self-evolution improves long-horizon embodied task success and memory QA scores over baseline controllers.
-
SOD: Step-wise On-policy Distillation for Small Language Model Agents
SOD reweights on-policy distillation strength step-by-step using divergence to stabilize tool use in small language model agents, yielding up to 20.86% gains and 26.13% on AIME 2025 for a 0.6B model.
-
CODA: Coordinating the Cerebrum and Cerebellum for a Dual-Brain Computer Use Agent with Decoupled Reinforcement Learning
A decoupled planner-executor GUI agent, trained by per-app reinforcement learning followed by specialist-to-generalist distillation, lifts ScienceBoard success from about 7.6% to 21.0% average and 40% pass@8.
-
SEAgent: Self-Evolving Computer Use Agent with Autonomous Learning from Experience
A self-evolving computer-use agent trained with full-trajectory state judging and curriculum task generation goes from 11.3% to 34.5% average success on five OSWorld apps, and a specialist-to-generalist variant beats ...
-
A Simple "Try Again" Can Elicit Multi-Turn LLM Reasoning
Multi-turn RL with only unary 'try again' feedback improves multi-turn reasoning accuracy by up to 14% while preserving single-turn performance.
-
How Far Can LLMs Improve from Experience? Measuring Test-Time Learning Ability in LLMs with Human Comparison
LLMs improve only slightly and unstably from test-time experience on semantic reasoning games, while humans learn much faster.
-
PAG: Multi-Turn Reinforced LLM Self-Correction with Policy as Generative Verifier
A new multi-turn reinforcement learning framework trains a single LLM to both solve math problems and verify its own solutions, revising only when its verifier finds a mistake.
-
Thinking vs. Doing: Agents that Reason by Scaling Test-Time Interaction
Scaling the number of interaction steps, trained via a curriculum over rollout horizon, improves web-agent task success and outperforms scaling per-step reasoning under fixed token budgets.
-
Self-Challenging Language Model Agents
A language model agent can generate its own verifiable training tasks and improve its tool-use success rate by about 2x without human-annotated data.
-
ARIA: Training Language Agents with Intention-Driven Reward Aggregation
Clustering language-agent actions into shared intentions and averaging their rewards reduces reward variance and improves policy performance in open-ended dialogue tasks.
-
Process Reward Models for LLM Agents: Practical Framework and Directions
A lightweight actor-critic loop trains 3B LLM agents to surpass GPT-4o on ALFWorld by learning step-level reward models from rollouts or demonstrations.
-
Certifiable Safe RLHF: Semantic Grounding and Fixed Penalty Constraint Optimization for Safer LLM Alignment
A fixed ReLU penalty and a semantically labeled cost model are proposed to make RLHF safer, but the 'certifiable' guarantee is not fully supported.
-
Automated Skill Discovery for Language Agents through Exploration and Iterative Feedback
EXIF repeatedly has a teacher agent explore an environment, relabel the exploration as tasks, train a student agent on it, and use the student's failures to guide the next round, improving 7B-8B agents in Webshop and Crafter.
-
Modeling and Optimizing User Preferences in AI Copilots: A Comprehensive Survey and Taxonomy
AI copilot preference optimization is organized into a pre-, mid-, and post-interaction taxonomy, with a unified definition of AI copilots.
-
SPA-RL: Reinforcing LLM Agents via Stepwise Progress Attribution
SPA-RL attributes a single final task reward to individual agent steps and uses the attributions plus an executability signal as dense rewards for PPO, yielding modest success-rate gains on WebShop, ALFWorld, and VirtualHome.
-
Response-Level Rewards Are All You Need for Online Reinforcement Learning in LLMs: A Mathematical Perspective
Under the assumption that the response reward equals the discounted sum of token rewards, response-level rewards suffice for unbiased token-level policy gradients in LLMs.
Discussion (0). Sign in to comment.