LLMInertia: A New Method to Improve Evidence Faithfulness of Large Language Models via Adaptive Anti-Inertia Reasoning

The Tsinghua University Machine Learning Group (THUMLP) proposed the LLMInertia method at ICML 2026, which addresses the "inertial thinking" problem of large language models (LLMs) through an adaptive anti-inertia reasoning mechanism, significantly improving evidence faithfulness and reasoning reliability. The related results have been open-sourced on GitHub (link: https://github.com/THUMLP/LLMInertia), released on 2026-06-03.

LLMs tend to exhibit the "inertial thinking" phenomenon when generating answers—after forming an initial judgment, subsequent reasoning will unconsciously seek evidence supporting that judgment, ignoring or downplaying contradictory evidence, leading to outputs deviating from factual basis and impairing evidence faithfulness (a core indicator for measuring LLM reliability). Existing studies show that even the most advanced LLMs cannot completely avoid this cognitive bias in complex reasoning tasks.

The core of LLMInertia is the adaptive anti-inertia reasoning mechanism: when a cognitive bias is detected in the reasoning steps, it automatically triggers the anti-inertia process, actively seeks neglected counter-evidence, and re-evaluates the rationality of the conclusion. Key components include: 1. Inertia detection module (identifies inertia nodes by analyzing changes in attention distribution and confidence fluctuations); 2. Evidence rebalancing mechanism (explicitly lists conflicting evidence and evaluates its credibility, simulating the "devil's advocate" strategy); 3. Adaptive fusion module (integrates original and anti-inertia reasoning results based on uncertainty weighting). This method can dynamically adjust the intensity and frequency of anti-inertia, balancing correction effect and reasoning efficiency.

LLMInertia has achieved significant improvements in multiple tasks: evidence faithfulness in fact-checking tasks increased by more than 15%; stable improvements were also observed in Science QA and multi-hop reasoning tasks. Meanwhile, anti-inertia reasoning is only triggered when there is a high risk of inertia, and the average increase in reasoning time is controlled within 20-30%, which is acceptable. The GitHub repository provides complete implementation and evaluation scripts to facilitate the reproduction of experimental results.

LLMInertia has theoretical and practical value: theoretically, it provides a new perspective for LLM reasoning improvement; practically, it can be applied to scenarios such as medical diagnosis assistance (avoiding premature diagnosis locking), legal document analysis (objectively evaluating all evidence), academic literature review (reducing confirmation bias), etc. Its adaptive feature can flexibly adjust the trigger threshold, adapt to resource-constrained environments, and balance evidence faithfulness and reasoning efficiency.

LLMInertia effectively mitigates cognitive biases in LLM reasoning and improves evidence faithfulness without significantly increasing computational overhead. As LLMs are increasingly applied in key fields, ensuring that outputs are faithful to evidence has become a core issue in AI research. LLMInertia provides a feasible technical path, and its open-source implementation lays the foundation for further research and application by the community. We look forward to more similar technologies emerging and being deployed in the real world in the future.