LLMInertia: Adaptive Counter-Inertial Reasoning to Improve Evidence Faithfulness in Large Language Models

This repository contains the implementation for the ICML 2026 paper:
"LLMInertia: Adaptive Counter-Inertial Reasoning to Improve Evidence Faithfulness in Large Language Models".

---

Requirements

The code is based on the LlamaFactory framework.

Induction Experiments

Train the Model with Inductive Data

bash examples/train_full/batch_train_multi_model_full.sh

• You can set the base model, base model template, and dataset to batch save model results.
• Remember to configure the yaml and save the model name.
• Add the dataset information in data/dataset_info.json.

---

Inference and Test Results

• Inference with the original (non-induced) model:

  bash evaluation/batch_vllm_infer_origin.sh

• Inference with the induced models:

  bash evaluation/batch_infer_eval_savesteps_full.sh

Hallucination Mitigation

python scripts/vllm_infer_evidence_medical.py