Why does majority voting outperform more complex inference methods?
Simple majority voting across independent samples often matches or beats sophisticated alternatives like Best-of-N and sequential revision. What makes this basic approach so hard to beat for reasoning models?
For reasoning models, majority voting across independent samples is a surprisingly strong baseline that sophisticated inference-time methods struggle to beat. Think Deep, Think Fast finds it generally competitive with or outperforming Best-of-N (which requires an external reward model) and sequential revision methods (which require the model to self-evaluate).
The robustness comes from what majority voting doesn't do: it doesn't require a verifier (which can be wrong), it doesn't require self-assessment (which reasoning models are poor at), and it doesn't rely on trace length (which is negatively correlated with correctness). It just exploits statistical redundancy across independent samples.
This doesn't mean majority voting is optimal — it's a ceiling-limited strategy. But it's the right default: simple, interpretable, and hard to beat without investing significantly in verifier quality. The research implication is that gains from more complex methods should be benchmarked against majority voting, not against single-sample baselines. Many reported improvements in the literature may not survive this comparison.
Extreme decomposition + voting at million-step scale (MAKER): The MAKER framework pushes majority voting to its logical extreme by decomposing complex tasks into atomic subtasks executed by microagents, each validated by voting. At scale (1000+ steps), this achieves error-free execution that no single-agent approach matches. MAKER also reveals scaling laws for multi-agent systems: more agents improve performance on complex tasks but hurt simple tasks (communication overhead exceeds benefit), and there's a critical complexity threshold below which single agents dominate. This extends the majority-voting baseline finding: voting's robustness is not just a property of independent sampling at the problem level — it works at every level of decomposition, from whole-problem voting down to atomic-subtask voting. The practical implication: when individual subtask accuracy is high (>95%), voting over decomposed subtasks compounds reliability multiplicatively. See Can extreme task decomposition enable reliable execution at million-step scale?.
Inquiring lines that read this note 15
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
Can validator consensus certify semantic correctness beyond agreement? Can preference-based training achieve better behavior optimization than supervised fine-tuning alone?- How does training-time voting differ from inference-time majority voting over samples?
- How does training-time consensus differ from inference-time majority voting over samples?
- What intermediate information does majority voting discard from reasoning chains?
- When does sequential reasoning provide exponential advantages over parallel voting?
- When does sequential chain-of-thought dramatically beat parallel voting approaches?
- What compute costs does majority-vote consensus sampling add versus supervised training?
Related concepts in this collection 9
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Why does parallel reasoning outperform single chain thinking?
Does dividing a fixed token budget across multiple independent reasoning paths beat spending it all on one long chain? This explores how breadth and diversity in reasoning compare to depth.
majority voting is the aggregation mechanism for parallel thinking
-
Does self-revision actually improve reasoning in language models?
When o1-like models revise their own reasoning through tokens like 'Wait' or 'Alternatively', does this reflection catch and fix errors, or does it introduce new mistakes? This matters because self-revision is marketed as a key capability.
why sequential revision methods underperform
-
Does voting discard useful reasoning from losing chains?
When multiple reasoning chains compete through majority voting, intermediate steps from non-winning chains are discarded. Could extracting and mixing those intermediate facts improve both the final answer and our ability to understand the reasoning?
extends: MCR shows voting is the correct baseline but not the ceiling; meta-reasoning over intermediate steps from all chains recovers distributed information that voting discards
-
Can extreme task decomposition enable reliable execution at million-step scale?
Can breaking tasks into maximally atomic subtasks with voting-based error correction solve the fundamental reliability problem in long-horizon tasks? This challenges whether better models or better decomposition is the path to high-reliability AI systems.
extends: voting works not just at problem level but at every decomposition level; MAKER scaling laws identify when multi-agent voting helps vs hurts
-
Can models trained on many imperfect experts outperform everyone?
Can generative models trained on diverse, biased experts achieve better performance than any individual contributor? This explores whether aggregating diverse perspectives during training acts as implicit denoising.
training-time analog: inference-time majority voting over samples from one model parallels the implicit majority vote over diverse training experts encoded in model weights
-
Can intermediate reasoning points yield better answers than final ones?
When reasoning models commit to a single path, they may miss better conclusions available at earlier decision points. Can aggregating completions from intermediate reasoning states recover lost accuracy?
sharpens the ceiling: voting at the final-answer level discards the intermediate-reasoning information that subthought aggregation extracts; aggregating modes from intermediate reasoning points within a single chain recovers up to 13% accuracy that final-answer voting cannot reach
-
Does self-consistency reliably reward correct answers during training?
Self-consistency initially correlates with correctness, but as models train on this signal, do they eventually learn to maximize consistency itself rather than accuracy? When does this proxy reward stop working?
names a structural risk in voting: when used as reward signal not just aggregation, the same statistical-redundancy property that makes voting robust also concentrates probability on consistent-but-wrong answers; voting's robustness is conditional on its use as aggregation rather than training signal
-
Does reward hacking always stem from the same failure?
Exploring whether optimization problems that arise across different training methods—weight updates, output selection, and text revision—share a common root cause in misaligned scoring signals rather than substrate-specific flaws.
places best-of-n selection among three substrates where optimizing against a scorer can find its blind spots; this note reports accuracy against Best-of-N and does not test for that exposure
-
How does the substrate change which behaviors an optimizer can reach?
Different optimization substrates—weights, selection, text—differ in which behaviors they can produce, what information they preserve, and what can be inspected or constrained. Understanding these differences matters because defenses or risk estimates may not transfer between substrates.
reads this note and the self-consistency note above as one agreement signal on two substrates: safe to select with, gamed when trained on; the vault's reading, not either source's
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- Deep Think with Confidence
- Think Deep, Think Fast: Investigating Efficiency of Verifier-free Inference-time-scaling Methods
- Can Large Language Models Capture Human Annotator Disagreements?
- Transcendence: Generative Models Can Outperform The Experts That Train Them
- Psychologically Enhanced AI Agents
- A Survey on Test-Time Scaling in Large Language Models: What, How, Where, and How Well?
- Let Me Think! A Long Chain-of-Thought Can Be Worth Exponentially Many Short Ones
- Reasoning Strategies in Large Language Models: Can They Follow, Prefer, and Optimize?
Original note title
majority voting is more robust than best-of-n and sequential revisions for reasoning models