Entity · paper

Sample More, Reflect Less: Self-Refine and Reflexion Lose to Repeated Sampling at Equal Token Cost, from 1.5B to 7B

paperactiveprovisional

sample-more-reflect-less-self-refine-and-reflexion-lose-to-repeated-sampling-at-equal-token-cost-from-1-5b-to-7b-ae9ea75f

·1 events·first seen 2h ago

Aliases: Sample More, Reflect Less: Self-Refine and Reflexion Lose to Repeated Sampling at Equal Token Cost, from 1.5B to 7B

Co-occurring entities

Reflexion Wang et al. 2024 Self-Refine Best-of-N

More like this (12)

On-Policy Self-Distillation with Sampled Demonstrations Reduces Output Diversity Self-Refine LESS: Mutual-Stability Sampling for Diffusion Language Models Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops Min-p sampling Partition, Prompt, Aggregate: Statistical Self-Consistency in Language Models Selective Timestep Weighting and Advantage-Based Replay for Sample-Efficient Diffusion RLHF CORE (Contrastive Reflection)Any-Dimensional Learning by Sampling Graph Sparse Sampling Copy Less, Ground More: Overcoming Repetitive Copying in Long-Context Reasoning via Evidence-Aware Reinforcement Learning intra-frame token sparsification

Recent events (1)

7arXiv · cs.CL·2h ago·source ↗

Rigorous experiment finds Self-Refine and Reflexion lose to repeated sampling at equal token cost across 1.5B–7B models

A new arXiv paper conducts a controlled experiment comparing seven LLM self-improvement methods (including Self-Refine, Reflexion, Best-of-N, and debate) against simple repeated sampling at equal token budgets, across 1.5B, 3B, and 7B open models on two math benchmarks. Using bootstrap confidence intervals and multiplicity correction across 36 paired comparisons, no method reliably outperforms repeated sampling; ten methods are reliably worse, all involving self-inspection of the model's own output. A notable finding is that Self-Refine and Reflexion remain 3.6–10.1 points below baseline even at 7B, and Reflexion on the smallest model silently degraded to a single chain-of-thought by always judging itself correct. The results challenge a broad class of iterative self-critique methods and extend earlier point-estimate findings by Wang et al. (2024) with proper statistical rigor.

Evaluation and Benchmarking Inference Economics Reflexion Wang et al. 2024 Self-Refine +3 more