DualStake: Dual-Path Confidence Calibration in Deep Research Agents
cs.CL, cs.AI, cs.LG
Submitted: 2026-09-01
Updated: 2026-09-01
Comments: Accepted to EMNLP 2026 Main
Code: https://github.com/FloXXXt/DualStake
License: http://creativecommons.org/licenses/by/4.0/
The gist: Deep Research agents tackle knowledge-intensive tasks through multi-round retrieval and decision-oriented generation.
Terminology
Abstract
Deep Research agents tackle knowledge-intensive tasks through multi-round retrieval and decision-oriented generation. However, these agents suffer from severe overconfidence, making their expressed confidence unreliable for user trust and downstream abstention. To address this, we augment the Deep Research pipeline with step confidence elicitation after each retrieval, building on the commonly used post-answer verbalized confidence. Interestingly, we find that Evidence Confidence (E-Conf), elicited after the final retrieval step, provides a stronger uncertainty signal than Answer Confidence (A-Conf), elicited after answer generation, and that A-Conf is largely shaped by E-Conf. Based on these findings, we propose DualStake, a dual-path calibration method that applies margin-clipped, confidence-dependent stake rewards to jointly align E-Conf and A-Conf with answer correctness while limiting extreme confidence optimization. Experiments on Qwen2.5-7B, Qwen2.5-7B-Instruct, and Qwen3-4B across 8 QA benchmarks demonstrate that DualStake consistently improves calibration without sacrificing answer accuracy. The code is available at https://github.com/FloXXXt/DualStake.
Sources
- Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty
- Language Models (Mostly) Know What They Know
- How do LLMs Compute Verbal Confidence
- NOVA: NOise-aware Verbal Confidence CAlibration for Robust Large Language Models in RAG Systems
- DeepResearch-Slice: Bridging the Retrieval-Utilization Gap via Explicit Text Slicing
- WebGPT: Browser-assisted question-answering with human feedback
- ZeroSearch: Incentivize the Search Capability of LLMs without Searching
- Text Embeddings by Weakly-Supervised Contrastive Pre-training
- A Comprehensive Survey on Trustworthiness in Reasoning with Large Language Models
- Measuring short-form factuality in large language models
- BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agents
- How to Train Your Deep Research Agent? Prompt, Reward, and Policy Optimization in Search-R1
- DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models
- The Confidence Dichotomy: Analyzing and Mitigating Miscalibration in Tool-Use Agents
- R1-Searcher: Incentivizing the Search Capability in LLMs via Reinforcement Learning
- Agentic Confidence Calibration
- Qwen3 Technical Report
Related papers
- Exploring Solution Divergence and Its Effect on Large Language Model Problem Solving
- Ishigaki-IDS-Bench: A Benchmark for Generating Information Delivery Specification from BIM Information Requirements
- Subliminal Steering: Stronger Encoding of Hidden Signals
- MedStruct-S: A Benchmark for Key Discovery, Key-Conditioned QA and Semi-Structured Extraction from OCR Clinical Reports
- The End of Transformers? On Challenging Attention and the Rise of Sub-Quadratic Architectures
- Untangling the Mechanisms of Misleading Context in Medical Question Answering