David Bai

A couple weeks ago, I participated in the Mercor x Etched x Cognition Hackathon. The theme of the hackathon was "inference-time compute"— and this is what I worked on for 24 hours, with some added rigor, visualizations, and analysis. You can view the original work here and the code here— my contribution is the hallucinations section for both. Thanks to Allison Lim and Vijay Kumaravelrajan ( alphabetical order) for helping me edit this! TL;DR: We can scale inference-time compute to gain signal for model uncertainty by injecting an interruption token, "No, but" at varying positions in a model's chain of thought. Sampling in parallel and measuring diversity of the reasoning traces allows us to then determine when a model should refuse to answer a question— that is, acknowledge it doesn't have the knowledge to accurately respond. We demonstrate improved refusal rates on questions a model would typically get wrong, while preserving accuracy on attempted questions, though there exists a mino

Interruption is All You Need: Reducing LLM Hallucination through Parallel Reasoning Diversity David Bai · March 23, 2025 · 10 minutes A couple weeks ago, I participated in the Mercor x Etched x Cognition Hackathon. The theme of the hackathon was "inference-time compute"— and this is what I worked on for 24 hours, with some added rigor, visualizations, and analysis. You can view the original work here and the code here — my contribution is the hallucinations section for both. Thanks to Allison Lim and Vijay Kumaravelrajan ( alphabetical order) for helping me edit this! TL;DR: We can scale infer

related reading

• Tracing the thoughts of a large language model \ Anthropicanthropic.com
• the-illusion-of-thinking.pdfml-site.cdn-apple.com
• Reasoning models don't always say what they think \ Anthropicanthropic.com
• DeepSeek-R1arxiv.org
• Thought Branches: Interpreting LLM Reasoning Requires Resamplingarxiv.org
• Announcing ReasoningLens — Visualizing and Diagnosing LLM Reasoning at a Glancehuggingface.co
• Reasoning as Trajectoriesslhleosun.github.io
• o1 and Reasoning | AndoLogsblog.ando.ai
• Features as Rewards: Using Interpretability to Reduce Hallucinationsgoodfire.ai
• The State of LLM Reasoning Model Inferencesebastianraschka.com
• Do reasoning models use their scratchpad like we do? Evidence from distilling paraphrasesalignment.anthropic.com
• Unfamiliar Finetuning Examples Control How Language Models Hallucinatearxiv.org