Research Papers 论文研究 2d ago Updated 1d ago 更新于 1天前 46

Selection, Recombination, or a Fresh Solve? A Candidate-Free Control for Single-Pass Test-Time Aggregation 选择、重组还是重新求解?一种用于单次测试时聚合的无候选者控制方法

Introduces a candidate-free control for single-pass test-time aggregation, comparing candidate conditioning against a fresh solve under identical token budgets On AIME-2025 and HMMT-2025 with Qwen3-4B, candidate conditioning boosts accuracy by +0.290 when multiple candidates are correct but reduces it by -0.123 when all candidates are wrong The one-correct candidate regime remains unresolved; the all-wrong recovery interpretation reverses under conservative correction for the adaptive two-benchm 候选聚合策略的价值高度依赖正确候选数量:多正确时显著提升准确率,全错误时反而降低 引入无候选控制实验,证明全错场景下"重新求解"优于"候选条件引导"(Δ_cand(c0) = -0.123) 在Qwen3-4B上验证:c2+提升+0.290,c0下降-0.123,c1结果未确定,结论经保守修正后仍成立 显式答案字段能因果引导模型输出朝向候选值,但掩码格式未带来可测量的准确率提升 研究局限于单一模型规模(4B)、两个数学基准、首答案截断候选片段和单遍聚合设置

62
Hot 热度
72
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • Introduces a candidate-free control for single-pass test-time aggregation, comparing candidate conditioning against a fresh solve under identical token budgets
  • On AIME-2025 and HMMT-2025 with Qwen3-4B, candidate conditioning boosts accuracy by +0.290 when multiple candidates are correct but reduces it by -0.123 when all candidates are wrong
  • The one-correct candidate regime remains unresolved; the all-wrong recovery interpretation reverses under conservative correction for the adaptive two-benchmark procedure
  • Structured interventions show explicit answer fields causally steer outputs, while masking yields no measurable accuracy gain
  • Evidence is scoped to one model family (Qwen3-4B), two math benchmarks, first-answer-truncated candidates, and single-pass aggregation prompts

Why It Matters

This work directly addresses a core tension in test-time compute scaling: whether conditioning an aggregation call on generated candidates provides genuine reasoning value or merely wastes tokens on redundant context. For practitioners building test-time reasoning pipelines, the findings suggest that candidate pools are a double-edged sword—helpful when diversity yields correct answers but harmful when all candidates share the same failure mode. The candidate-free control methodology also sets a rigorous benchmark for evaluating future aggregation strategies.

Technical Details

  • Candidate-free control: The paper introduces a baseline where the aggregation model solves the problem afresh without any candidate context, matched to the same maximum output-token allowance as the candidate-conditioned variant
  • Stratification by correct candidate count: Results are split into three regimes—c0 (all wrong), c1 (one correct), and c2+ (multiple correct)—revealing asymmetric effects of candidate conditioning
  • Benchmarks and model: Evaluated on AIME-2025 and HMMT-2025 mathematics competitions using Qwen3-4B, with first-answer-truncated candidate fragments and single-pass prompted aggregation
  • Structured intervention: Explicit answer fields were shown to causally steer model outputs toward their values; masking interventions produced no measurable accuracy improvement, and format equivalence was not established
  • Statistical correction: Conclusions for c2+ and c0 survive a conservative correction for the adaptive two-benchmark procedure, strengthening confidence in those findings

Industry Insight

  • Test-time aggregation strategies should be regime-aware: invest in candidate generation and conditioning when expecting multiple plausible paths, but consider fresh solves when candidates are likely uniformly incorrect
  • The negative effect of all-wrong candidate conditioning (-0.123) suggests that polluted candidate context can actively mislead aggregation models, implying that candidate filtering or quality gates may be essential before aggregation
  • Explicit output formatting (e.g., structured answer fields) is a leveraged causal intervention worth incorporating into pipeline design, while naive masking approaches should be abandoned in favor of format-aware prompting

TL;DR

  • 候选聚合策略的价值高度依赖正确候选数量:多正确时显著提升准确率,全错误时反而降低
  • 引入无候选控制实验,证明全错场景下"重新求解"优于"候选条件引导"(Δ_cand(c0) = -0.123)
  • 在Qwen3-4B上验证:c2+提升+0.290,c0下降-0.123,c1结果未确定,结论经保守修正后仍成立
  • 显式答案字段能因果引导模型输出朝向候选值,但掩码格式未带来可测量的准确率提升
  • 研究局限于单一模型规模(4B)、两个数学基准、首答案截断候选片段和单遍聚合设置

为什么值得看

这篇论文为测试时推理中的候选聚合策略提供了关键的对照实验证据,揭示了"候选上下文"在不同正确率场景下的差异化价值。对AI从业者而言,这有助于优化测试时计算资源的分配策略——在候选全错时避免盲目依赖候选信息,转而采用重新求解策略。

技术解析

  • 实验设计:在AIME-2025和HMMT-2025数学基准上,使用Qwen3-4B模型,按正确候选数量(c0/c1/c2+)分层评估候选条件 vs 无候选控制的准确率差异,保持相同最大输出token限制
  • 核心发现:c2+时Δ_cand=+0.290(候选条件显著提升),c0时Δ_cand=-0.123(候选条件反而降低准确率),c1结果未确定;结论在针对自适应双基准程序的保守修正后仍然成立
  • 结构化干预实验:显式答案字段(如"Answer: ___")能因果引导模型输出朝向候选值;掩码格式未带来可测量的准确率提升,且与原始格式不等价
  • 局限性:证据仅来自Qwen3-4B家族、两个数学基准、首答案截断的候选片段和单遍提示聚合

行业启示

  • 测试时推理策略应动态适配:当候选质量高时利用候选聚合,全错时优先重新求解而非强行聚合,避免候选污染
  • 提示工程中的格式设计(如显式答案字段)具有因果影响力,可被用于引导输出,但需验证格式等价性,不能假设不同格式效果相同
  • 当前结论局限于小模型(4B)和数学领域,推广至更大模型或更复杂任务(如代码、开放域)需谨慎,建议在不同规模模型和任务上验证

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Inference 推理 Research 科学研究 Evaluation 评测