Sample More, Reflect Less: Self-Refine and Reflexion Lose to Repeated Sampling at Equal Token Cost

Forum topic · ✨步子哥 · 2026-07-31

Summary

A forum post discusses a paper (arXiv: 2607.28576) showing that at equal token budgets, repeated sampling with majority voting consistently beats iterative self-refinement methods like Self-Refine and Reflexion across models from 1.5B to 7B on benchmarks such as GSM8K, GPQA, and MATH. The post argues that earlier comparisons were unfair because Self-Refine and Reflexion consume multiple rounds of tokens while single-pass generation does not. Once budgets are equalized, refinement's apparent gains largely vanish: models often 'improve' correct answers into worse ones or merely revert to mediocre versions, wasting tokens on noisy self-critique. The finding scales across model sizes, suggesting current LLMs' self-correction ability is far weaker than assumed. Practical advice: prefer repeated sampling plus majority voting, and always compare refinement pipelines against an equal-token sampling baseline. Limitations include the narrow 1.5B–7B model range and comparison by token cost rather than latency, so results may differ for frontier models.

This forum post discusses the paper Sample More, Reflect Less: Self-Refine and Reflexion Lose to Repeated Sampling at Equal Token Cost, from 1.5B to 7B (arXiv: 2607.28576).

The Core Question

Self-Refine and Reflexion—two of the most cited "reasoning enhancement" methods—work by having an LLM generate an answer, critique itself, and rewrite iteratively. The paper asks a simple question: at an equal token budget, does this reflective iteration actually beat simply sampling multiple independent answers and voting?

The answer is no. From 1.5B to 7B models, across six or seven benchmarks, repeated sampling consistently beats Self-Refine and Reflexion when token budgets are equalized.

Why Equal Token Budgets Matter

Earlier literature claiming "Self-Refine beats direct generation" contained a hidden confound: Self-Refine consumes multiple rounds of tokens while direct generation uses only one. The paper controls for this by giving both methods the same token budget. Results:

On GSM8K, GPQA, MATH and other reasoning tasks, repeated sampling consistently wins.
Larger models (7B vs 1.5B) show an even bigger sampling advantage—higher per-sample quality makes it easier to hit correct answers.
Refinement's "improvements" often just rewrite an already-correct answer into a more mediocre version—reflection smooths rather than corrects.

What Reflection Actually Does

A sharp observation: much of Self-Refine/Reflexion's "improvement" happens on cases where the original answer was already correct—the model critiques itself, turns right answers wrong, then sometimes back again. This back-and-forth burns tokens for near-zero or negative net gain.

The core problem is that reflection methods don't know when to revise and when not to. They force a critique-rewrite loop on every problem, and the model's self-critique is as error-prone as its generation. Repeated sampling, by contrast, requires no self-knowledge—only that the model occasionally guesses correctly. Letting the model guess multiple times and picking the best is more reliable than having it find fault with itself.

Why It Matters

Over the past two years, "letting models reflect" became almost a doctrine. This paper suggests some of these methods' gains may come not from reflection itself but from spending several times more tokens. Once budgets are equalized, the halo fades. The "from 1.5B to 7B" in the title signals this is scale-independent, hinting that current LLMs' self-critique ability may be far weaker than assumed: a model critiquing itself shares the same biases and blind spots as when it generates.

Practical Advice

1. Prefer repeated sampling + majority voting over Self-Refine or Reflexion. 2. Sampling is simpler at equal budget—no critique prompts, no multi-turn dialogue management, just sample N times and take the mode. 3. If you must use reflection, benchmark against an equal-token sampling baseline first—your reflection pipeline may underperform a one-line for i in range(N) loop.

Honest Limitations

Only 1.5B–7B models were tested; whether reflection wins at 70B+ or GPT-4-class scale is unknown.
Comparison is by token budget, not wall-clock time or money. Sampling is naturally parallel while reflection is serial—reflection may still have value in latency-sensitive settings.

But the core conclusion stands firm: under the most common comparison baseline—token budget—reflection loses to sampling. Earlier "Self-Refine significantly improves performance" papers owe the community an equal-token controlled experiment.

---

Paper link: https://arxiv.org/abs/2607.28576

Related code: Self-Refine original repository https://github.com/madaan/self-refine (the paper itself provides no independent code repository)