Research map / Large language models / Reasoning and chain of thought
Truthfulness Does Not Scale Like Reasoning: Why Polling Fails as a Proxy Verifier
Yegor Denisov-Blanch, Joshua Kazdan, Jessica Chudnovsky, Rylan Schaeffer, Sheng Guan, Soji Adeshina, Sanmi Koyejo
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Examines whether scaling inference compute via polling (sampling multiple answers and aggregating) improves truthfulness in domains lacking external verifiers, finding it does not scale like reasoning gains.
Problem
Whether inference-time compute scaling (e.g., pass@k style sampling and aggregation) can elicit truthfulness gains in domains without convenient external verifiers, unlike math and code.
Approach
Tests polling-based inference scaling as a proxy verifier for truthfulness across five benchmarks and models, and argues that truthfulness does not scale like reasoning with more compute.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| aggregate samples by majority vote instead of a single sample | LLM ensembles · HLE, BoolQ, Com2Sense, Predict-the-Future | accuracy | No clear effect |
| scale to 25x inference cost with polling aggregation | LLMs · verifier-absent benchmarks | accuracy | No clear effect |
| weight votes by self-reported confidence | majority voting · HLE | accuracy (%): 5.7 -> 2.9 | Hurt |
| weight votes by self-reported confidence | majority voting · Com2Sense | accuracy (%): 85.0 -> 82.0 | Hurt |
| use highest-confidence answer | majority voting · HLE | accuracy (%): 5.7 -> 11.4 | Improved |
| use Surprisingly Popular rule | majority voting · HLE | accuracy (%): 5.7 -> 20.0 | Improved |
| use Surprisingly Popular rule | majority voting · BoolQ | accuracy (%): 77.0 -> 76.0 | Hurt |
| invert the SP signal | SP aggregation · HLE | accuracy: 80% | Improved |
| use inverse-SP | SP aggregation · other datasets | accuracy | No clear effect |
| aggregate on forecasting questions postdating knowledge cutoffs | all aggregation methods · Predict-the-Future | accuracy | No clear effect |
| raise temperature from 0.7 to 1.0 to induce diversity (ablation) | LLM sampling · question-model pairs | plurality answer flips: 2.9% | No clear effect |
| pool across five models instead of one model (ablation) | ensembles · MATH | convergence to same wrong answer: 53% | Hurt |
Surprising findings
- Scaling inference compute via polling does not yield the truthfulness gains seen for reasoning tasks with external verifiers.
Experimental setup
Builds on: pass@k
Closest papers by meaning
- The Limits of Inference Scaling Through Resampling (ICLR 2026 · similarity 0.88)
- Scaling Inference-Efficient Language Models (ICML 2025 · similarity 0.88)
- To Trust or Not to Trust? Enhancing Large Language Models' Situated Faithfulness to External Contexts (ICLR 2025 · similarity 0.88)
- A Theoretical Study on Bridging Internal Probability and Self-Consistency for LLM Reasoning (NeurIPS 2025 · similarity 0.88)
- Is Best-of-N the Best of Them? Coverage, Scaling, and Optimality in Inference-Time Alignment (ICML 2025 · similarity 0.88)
- Beyond Model Collapse: Scaling Up with Synthesized Data Requires Verification (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
