Research map / Multi-agent systems and games / LLM agents

Uncertainty-Aware Clarification in LLM Agents with Information Gain

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Trains a small LLM clarifier with DAPO using an information-gain reward (shift in length-normalized teacher-forced log-likelihood of the ground-truth goal) to ask targeted clarification questions in a clarification-augmented tau-Bench.

Problem

LLM agents act on underspecified user instructions, causing erroneous tool actions; existing training gives little signal on when and what clarification questions reduce uncertainty about user intent.

Approach

Reward clarification questions by the Bayesian belief update they induce toward the ground-truth goal (posterior minus prior teacher-forced log-likelihood, equivalent to a log importance weight/PMI), and optimize a clarifier policy with group-relative DAPO, using a strict user simulator to prevent reward hacking.

Key results

• Ours (Qwen3-1.7B clarifier) improves average success from 15.4% (None) to 17.8% across retail and airline, with average clarifier calls dropping from 4.2 (untrained baseline) to 1.3.
• Averaged over five agent backbones, the learned clarifier raises success from 23.6% to 27.3% (+3.7%), versus -0.6% for a Qwen3-8B clarifier.
• Without the information-gain term (likelihood-only), airline success falls from 17.3% to 10%; LLM-as-judge reward yields 8% on airline.

Surprising findings

• The interactive clarifier outperforms the full-user-intent oracle on the airline domain (17.3% vs 16%), which the authors attribute to avoiding information overload.
• A clarifier 8B in size (Qwen3-8B) causes performance regressions on several agents, showing general-purpose clarifiers can hurt.

Limitations

• Reward requires access to ground-truth user goals, available only in controlled environments.
• Base agent kept fixed; joint agent-clarifier optimization not explored.
• Clarification alone does not fix downstream execution/policy errors, as shown in failure cases.

Experimental setup

Datasets: tau-Bench retail trajectories tau-Bench airline tasks

Benchmarks: tau-Bench retail tau-Bench airline

Metrics: Pass@1 success rate steps clarification count

Baselines: No clarification (None) Qwen3-8B clarifier LLM as a Judge reward w/o Information Gain (likelihood-only) Qwen3-30B DeepSeek-R1 DeepSeek-V3.1 Full user intent oracle

Builds on: DAPO GRPO tau-Bench IGPO VERL

Closest papers by meaning

• Position: Uncertainty Quantification Needs Reassessment for Large Language Model Agents (ICML 2025 · similarity 0.90)
• Ask Early, Ask Late, Ask Right: When Does Clarification Timing Matter for Long-Horizon Agents? (NeurIPS 2026 · similarity 0.90)
• ECLAIR: Enhanced Clarification for Interactive Responses in an Enterprise AI Assistant (AAAI 2025 · similarity 0.89)
• ECLAIR: Enhanced Clarification for Interactive Responses (AAAI 2025 · similarity 0.89)
• Enhancing Decision-Making for LLM Agents via Step-Level Q-Value Models (AAAI 2025 · similarity 0.88)
• ASPI: Seeking Ambiguity Clarification Amplifies Prompt Injection Vulnerability in LLM Agents (NeurIPS 2026 · similarity 0.87)