ReHallucination Inference-Time Intervention

This repository contains the inference-time intervention code and released data for the paper "Answering the Unanswerable". The code focuses on the inference pipeline: vanilla trajectory generation, probe-based intervention, and evaluation.

Overview

The method is organized into three stages:

1. Stage 1: Vanilla generation Generate reasoning trajectories without intervention for answerable and unanswerable problems.

2. Stage 2: Probe detection and intervention Replay the vanilla trajectory, use a linear probe at wait positions to detect the first intervention point, insert the intervention prompt once, and continue generation.

3. Stage 3: Evaluation Evaluate unanswerable examples with Abstention and evaluate answerable examples with Answer Accuracy.

Precomputed Stage 1 and Stage 2 outputs are included under data/out/, so you can directly evaluate released results without rerunning generation.

Repository Layout

codes/
  main_vllm.py                  # Stage 1 vanilla generation
  a_inter.py                    # Stage 2 probe detection + intervention
  interven.py                   # Linear probe utilities
  stage_1_generate_vanilla.sh   # Stage 1 example script
  stage_2_run_intervention.sh   # Stage 2 example script
  stage_3_evaluate.py           # Stage 3 evaluator
  stage_3_evaluate.sh           # Stage 3 example script
  a_input_path.py               # Paths to released vanilla trajectories

data/
  sum/                          # SUM input data
  umwp/                         # UMWP input data
  out/stage1/                   # Released vanilla trajectories
  out/stage2/                   # Released intervention outputs

model/
  *_layer_result/layer_*.pt     # Released linear probe weights

Setup

Install the Python dependencies needed by the stages you want to run.

pip install -r requirements.txt

The default Stage 3 evaluator uses a lightweight exact/numeric matcher and does not require a local judge model.

Stage 1: Generate Vanilla Trajectories

Stage 1 generates answerable and unanswerable vanilla reasoning trajectories.

cd codes
bash stage_1_generate_vanilla.sh

Useful environment variables:

MODEL_PATH=deepseek-ai/DeepSeek-R1-Distill-Llama-8B
DATASET=SUM            # SUM or UMWP
RUN_SPLIT=both         # both, answer, or unanswer

Outputs are written to:

data/out/stage1/{DATASET}/{MODEL_NAME}/{solve,unsolve}/

Released Stage 1 outputs are already included and referenced by codes/a_input_path.py.

Stage 2: Run Intervention

Stage 2 first detects the intervention point with a linear probe, then performs one intervention by inserting the fixed intervention prompt and continuing generation.

Example:

cd codes
bash stage_2_run_intervention.sh

Default example configuration:

DATASET=SUM
MODEL_NAME=DeepSeek-R1-Distill-Llama-8B
MODEL_PATH=deepseek-ai/DeepSeek-R1-Distill-Llama-8B
PROBE_LAYER=22
THEROD=0.4

Outputs are written to:

data/out/stage2/{DATASET}/{MODEL_NAME}/
  unsolve_inter_q_id_point.json
  solve_inter_q_id_point.json
  unsolve_intervention_result.jsonl
  solve_intervention_result.jsonl

Released Stage 2 intervention outputs are included for SUM and UMWP across the released models.

Probe Layers

The released probe weights cover the layers used in our experiments:

Dataset	Model	Probe Layer
SUM	DeepSeek-R1-Distill-Llama-8B	22
SUM	DeepSeek-R1-Distill-Qwen-7B	17
SUM	DeepSeek-R1-Distill-Qwen-14B	30
SUM	Qwen3-8B	24
SUM	Qwen3-14B	26
UMWP	DeepSeek-R1-Distill-Llama-8B	17
UMWP	DeepSeek-R1-Distill-Qwen-7B	18
UMWP	DeepSeek-R1-Distill-Qwen-14B	30
UMWP	Qwen3-8B	20
UMWP	Qwen3-14B	24

Stage 3: Evaluate Results

To evaluate released Stage 2 outputs:

cd codes
bash stage_3_evaluate.sh

By default, this evaluates:

data/out/stage2/SUM/DeepSeek-R1-Distill-Llama-8B/

Change the target with environment variables:

DATASET=UMWP MODEL_NAME=Qwen3-8B bash stage_3_evaluate.sh

The evaluator reports:

• Abstention on unanswerable examples.
• Answer Acc on answerable examples.
• A JSON metrics file at data/out/stage2/{DATASET}/{MODEL_NAME}/stage3_metrics.json.

By default, answer equivalence uses a lightweight exact/numeric matcher:

ANSWER_JUDGE=exact bash stage_3_evaluate.sh

If you want to use an LLM judge for answer equivalence, configure your local judge service and pass the corresponding options to stage_3_evaluate.py.

Notes

The Stage 2 implementation uses vanilla trajectories from Stage 1, which approximates online intervention while making experiments reproducible.