← Về thư mục

DualStake: Dual-Path Confidence Calibration in Deep Research Agents

Accepted to: EMNLP 2026 Main Conference

Overview

Large language model agents equipped with search tools can collect evidence over multiple retrieval rounds, but their expressed confidence is often poorly aligned with answer correctness. We introduce DualStake, a confidence-calibration method for deep research agents.

Built on Search-R1, DualStake augments the agent trajectory with two confidence signals: Evidence Confidence, elicited after retrieval, and Answer Confidence, elicited after the final answer. It jointly supervises both signals using a margin-clipped stake reward, improving calibration while preserving answer quality.

Getting Started

Installation

This codebase follows the environment setup of Search-R1. After creating that environment, install the dependencies and this package:

pip install -r requirements.txt
pip install -e .

Data Preparation

Prepare Search-R1-compatible JSONL files containing question and golden_answers:

/path/to/raw-data/
├── nq/
│   ├── train.jsonl
│   └── test.jsonl
└── hotpotqa/
    ├── train.jsonl
    └── dev.jsonl

Build the NQ + HotpotQA training mixture used in the paper:

python scripts/build_data.py \
  --raw-root /path/to/raw-data \
  --output-dir data/nq_hotpotqa

Retrieval Service

Following Search-R1, build or obtain an E5 index over the 2018 Wikipedia corpus, then launch the local retrieval service:

INDEX_PATH=/path/to/e5_Flat.index \
CORPUS_PATH=/path/to/wiki-18.jsonl \
bash scripts/serve_retriever.sh

The default retrieval endpoint is http://127.0.0.1:8000/retrieve.

Training

The training script uses the paper's balanced DualStake setting: GRPO with 5 rollouts, top-3 E5 retrieval, (\alpha=\beta=0.25), linear warm-up from steps 100 to 300, and stake clipping to ([0.1, 0.9]).

DATA_DIR=./data/nq_hotpotqa \
MODEL_PATH=Qwen/Qwen2.5-7B \
RETRIEVER_URL=http://127.0.0.1:8000/retrieve \
NUM_GPUS=8 \
bash scripts/train_dualstake.sh

Evaluation

Evaluate a base model or trained checkpoint on a prepared benchmark split:

MODEL_PATH=/path/to/checkpoint \
DATA_DIR=/path/to/benchmark \
OUTPUT_DIR=./evaluation/benchmark \
RETRIEVER_URL=http://127.0.0.1:8000/retrieve \
bash scripts/evaluate.sh

Use scripts/compute_metrics_ext.py and scripts/calc_last_step_conf.py to calculate calibration metrics from generated evaluation statistics.

Acknowledgements

DualStake is built on Search-R1 and includes modified components from veRL. We thank their authors and contributors for making their work available. Please preserve the included LICENSE and Notice.txt when redistributing this code.