Critical-State RL: Diagnosing Trainable States for Multi-Turn Tool Use

TL;DR

Critical-State RL identifies trainable states in multi-turn tool use, improving task success by about 14 percentage points in missing-function tasks.

cs.LG 🔴 Advanced 2026-09-22 4 views
Zixiang Chen Wenting Zhao Zhepeng Cen Akshara Prabhakar Jielin Qiu Jianguo Zhang Zhiwei Liu Tulika Manoj Awalgaonkar Liangwei Yang Shelby Heinecke Silvio Savarese Huan Wang
reinforcement learning multi-turn interaction tool use state diagnosis contextual training

Key Findings

Methodology

Critical-State RL uses nested sampling to separate action-dependent reward variation from continuation noise, optimizing policies through contextual-bandit training. It identifies trainable states in multi-turn interactions by assessing whether each reward captures the action's effect on task success and if improvement over a reference policy is possible.

Key Results

  • On the Berkeley Function Calling Leaderboard (BFCL) v4, the diagnostic selects the response after tool availability for missing-function tasks, improving performance by about 14 percentage points.
  • For missing-argument tasks, it selects the response before the missing argument is supplied, improving performance by about 3.8 percentage points.
  • Training alternative responses does not improve performance and may even worsen it.

Significance

This research addresses the challenge of identifying which model calls in multi-turn interactions would benefit from training, providing a new approach to state selection in multi-turn tool use. It holds significant academic and industrial value, particularly in optimizing complex tasks.

Technical Contribution

Introduces a training-free diagnostic method that separates action-dependent reward variation from continuation noise, guiding local training. Compared to existing state selection methods, it offers new theoretical guarantees and engineering possibilities.

Novelty

First to use nested sampling to separate reward variation from noise in multi-turn interactions, diagnosing trainable states. This innovation provides a new perspective and tool compared to existing state selection methods.

Limitations

  • In some cases, the diagnostic may not accurately identify all critical states, especially in complex task environments.
  • The method may have high computational costs, particularly on large datasets.
  • Further research is needed to verify its generalizability across different tasks and models.

Future Work

Future work could explore the application of this method to more tasks and models, particularly in real-time systems. Additionally, researching ways to reduce computational costs is an important direction.

AI Executive Summary

Failures in multi-turn tool use often hinge on a single model call, yet reward variation alone does not reveal which calls would benefit from training. Critical-State RL addresses this by identifying trainable states in multi-turn interactions. The method uses nested sampling to separate action-dependent reward variation from continuation noise and optimizes policies at selected states using contextual-bandit training.

In experiments on the Berkeley Function Calling Leaderboard (BFCL) v4, Critical-State RL selects the response after tool availability for missing-function tasks, improving performance by about 14 percentage points. For missing-argument tasks, it selects the response before the missing argument is supplied, improving performance by about 3.8 percentage points. Training alternative responses does not improve performance and may even worsen it.

This research provides a new approach to state selection in multi-turn tool use, holding significant academic and industrial value. Future work could explore the application of this method to more tasks and models, particularly in real-time systems. Additionally, researching ways to reduce computational costs is an important direction.

Deep Analysis

Background

Multi-turn tool use involves a sequence of model decisions, yet trajectory-level rewards do not identify which decisions would benefit from training. Existing research focuses on turn-level credit assignment and state selection for local training, but these methods fail to address the challenge of identifying which model calls need training.

Core Problem

Failures in multi-turn tool use often hinge on a single model call, yet reward variation alone does not reveal which calls would benefit from training. The key challenge is identifying states that significantly impact task success.

Innovation

Critical-State RL uses nested sampling to separate action-dependent reward variation from continuation noise, diagnosing trainable states. This innovation allows for more accurate identification of states needing training in multi-turn interactions.

Methodology

  • �� Define task-specific candidate calls and local rewards.
  • �� Use nested sampling to separate action-dependent reward variation from continuation noise.
  • �� Optimize policies through contextual-bandit training.
  • �� Conduct local training at selected states.

Experiments

Experiments on the Berkeley Function Calling Leaderboard (BFCL) v4 compare training at diagnostic-selected states with training at alternative states. Missing-function and missing-argument tasks are tested, recording performance changes.

Results

For missing-function tasks, selecting the response after tool availability improves performance by about 14 percentage points. For missing-argument tasks, selecting the response before the missing argument is supplied improves performance by about 3.8 percentage points.

Applications

This method can be applied to state selection in multi-turn tool use, particularly in optimizing complex tasks. It helps identify states that significantly impact task success, improving model performance.

Limitations & Outlook

In some cases, the method may not accurately identify all critical states, especially in complex task environments. Additionally, the computational cost may be high, particularly on large datasets.

Plain Language Accessible to non-experts

Imagine you're cooking in a kitchen. You need to use the right tool at the right time, like not adding salt before boiling. Critical-State RL is like a smart assistant that tells you when to use each tool to ensure your dish turns out perfect. It watches each step's effect and tells you which step needs improvement, just like ensuring every step in cooking is flawless.

ELI14 Explained like you're 14

Imagine you're playing a game where you need to use tools to solve problems. Sometimes, you need to use the right tool at the right time. Critical-State RL is like a hint system in the game that tells you when to use each tool to ensure you pass the level. It watches your every move and tells you which step needs improvement, just like making sure every step in the game is perfect.

Glossary

Critical-State RL

A method for identifying trainable states in multi-turn interactions by separating action-dependent reward variation from continuation noise.

Used to diagnose trainable states in multi-turn tool use.

Nested Sampling

A technique for separating action-dependent reward variation from continuation noise.

Used in Critical-State RL to identify trainable states.

Contextual-Bandit Training

A training method for optimizing policies at selected states.

Used in Critical-State RL to optimize policies.

Berkeley Function Calling Leaderboard (BFCL)

A benchmark dataset for evaluating multi-turn tool use tasks.

Used to validate the effectiveness of Critical-State RL.

Action-dependent Reward Variation

The portion of reward variation caused by the current action.

Used in Critical-State RL to separate reward variation.

Open Questions Unanswered questions from this research

  • 1 How to more accurately identify critical states in complex task environments remains to be studied.
  • 2 Reducing computational costs for application on large datasets is an important direction.
  • 3 Exploring the method's performance in real-time systems needs further verification.

Applications

Immediate Applications

Multi-turn Tool Use Optimization

Helps identify states that significantly impact task success, improving model performance.

Long-term Vision

Real-time System Application

Explores the method's performance in real-time systems, particularly in complex task environments.

Abstract

Multi-turn tool-use failures can hinge on a single model call, yet reward variation alone does not reveal which call would benefit from training. When rewards depend on later interactions, their variation can reflect downstream randomness rather than differences between the current actions. We introduce Critical-State RL to identify trainable states in multi-turn interactions. Given task-defined candidate calls and local rewards, the method assesses whether each reward captures the action's effect on task success and whether improvement over a reference policy is possible. It then uses nested sampling to separate action-dependent reward variation from continuation noise and optimizes the policy at the selected states using contextual-bandit training. Experiments on the Berkeley Function Calling Leaderboard (BFCL) v4 compare training at diagnostic-selected states with training at alternative states. For missing-function tasks, the diagnostic selects the response after the tool becomes available; for missing-argument tasks, it selects the response before the missing argument is supplied. Training the selected responses improves performance, including about 14 percentage points on the missing-function task, while training the alternatives leaves performance flat or worse. We further apply the recipe across models and tasks, including logged repeat-call avoidance and memory management.

cs.LG cs.CL