Research Papers 论文研究 23h ago Updated 20h ago 更新于 20小时前 43

Mirror Horizon: Viable Path Entropy as a Measure of Bounded Reflection 镜像视界:有界反思的可行路径熵度量

Introduces Mirror Theory and Viable Path Entropy (VPE) as a metric for measuring bounded reflection and coherent continuation capacity in intelligent systems. VPE decomposes capability into the probability of reaching a viable continuation and the diversity of verified continuation modes, moving beyond simple accuracy metrics. Experiments on GSM8K with Qwen2.5-Instruct models show that increasing token budget from 96 to 160 significantly improves verified reachability and smoothed VPE. The Qwen2 提出“镜像理论”(Mirror Theory),主张智能系统的评估应关注其在重复反思下维持连贯延续的能力,而非仅看静态表征。 定义“可行路径熵”(VPE)作为有限预算下的验证延续能力度量,分解为到达可行延续的概率与成功 rollout 中验证模式的多样性。 在 GSM8K 数据集上的 Qwen2.5-Instruct 系列实验显示,增加 Token 预算(从 96 增至 160)显著提升了验证可达性和平滑 VPE。 发现 Qwen2.5-1.5B 在 160 Token 预算下展现出比参数更多的 Qwen2.5-3B 更强的“镜像视界”,证明能力取决于受限反射协议下的可用验证延续容量,而非单纯

55
Hot 热度
70
Quality 质量
60
Impact 影响力

Analysis 深度分析

TL;DR

  • Introduces Mirror Theory and Viable Path Entropy (VPE) as a metric for measuring bounded reflection and coherent continuation capacity in intelligent systems.
  • VPE decomposes capability into the probability of reaching a viable continuation and the diversity of verified continuation modes, moving beyond simple accuracy metrics.
  • Experiments on GSM8K with Qwen2.5-Instruct models show that increasing token budget from 96 to 160 significantly improves verified reachability and smoothed VPE.
  • The Qwen2.5-1.5B model achieved the strongest mirror horizon at 160 tokens, outperforming the larger 3B model, demonstrating that accessible verified continuation capacity matters more than parameter count.

Why It Matters

This research shifts the focus from static performance metrics like pass@k to dynamic measures of an AI's ability to sustain coherent reasoning over multiple reflection steps. For practitioners, it highlights that model efficiency and reasoning stability can be optimized through better reflection protocols rather than just scaling parameter counts. It offers a new theoretical framework for evaluating and improving the robustness of LLMs in complex reasoning tasks.

Technical Details

  • Viable Path Entropy (VPE): A finite-budget measure defined by four components: mirror state, rollout protocol, verifier, and mode map. It quantifies both the likelihood of success and the diversity of successful strategies.
  • Theoretical Scaffold: The paper defines intuition as local underdetermining constraints, taste as invariant-selecting pressure, reflection as taste-guided resolution, and geometry as the learned structure enabling stable future reflection.
  • Experimental Setup: Evaluated on GSM8K using Qwen2.5-Instruct models (1.5B and 3B). Each problem involved 32 sampled rollouts across two reflection horizons.
  • Key Findings: Increasing the token budget from 96 to 160 reduced zero-reachability and increased verified-mode entropy. The 1.5B model surpassed the 3B model in mirror horizon strength, proving that bounded reflection capacity is distinct from raw model size.

Industry Insight

  • Rethinking Model Evaluation: Developers should consider dynamic reasoning capabilities and continuation stability alongside traditional accuracy benchmarks when selecting models for complex, multi-step tasks.
  • Efficiency Over Scale: Smaller models can outperform larger ones in specific reasoning protocols if they possess higher verified continuation capacity. Optimizing reflection budgets and verifiers may yield better results than simply deploying larger models.
  • Protocol Design Importance: The effectiveness of AI systems is heavily dependent on the rollout and verification protocols used. Investing in better "taste" mechanisms and reflection horizons can unlock latent capabilities in existing models.

TL;DR

  • 提出“镜像理论”(Mirror Theory),主张智能系统的评估应关注其在重复反思下维持连贯延续的能力,而非仅看静态表征。
  • 定义“可行路径熵”(VPE)作为有限预算下的验证延续能力度量,分解为到达可行延续的概率与成功 rollout 中验证模式的多样性。
  • 在 GSM8K 数据集上的 Qwen2.5-Instruct 系列实验显示,增加 Token 预算(从 96 增至 160)显著提升了验证可达性和平滑 VPE。
  • 发现 Qwen2.5-1.5B 在 160 Token 预算下展现出比参数更多的 Qwen2.5-3B 更强的“镜像视界”,证明能力取决于受限反射协议下的可用验证延续容量,而非单纯参数量。

为什么值得看

该研究挑战了以单次准确率或 Pass@k 为主的传统大模型评估范式,引入了基于动态反思和延续能力的评估维度。对于 AI 从业者而言,这提供了理解小模型为何能在特定推理协议下超越大模型的新的理论视角,即“结构化的可行延续”比“参数规模”更能决定复杂推理中的稳定性。

技术解析

  • 核心理论框架:镜像理论将直觉视为局部欠确定约束,品味作为不变量选择压力,反思作为品味引导的欠确定性解决机制,几何结构则是使未来反思稳定的习得结构。
  • VPE 度量指标:可行路径熵由两部分组成:一是达到可行延续的概率,二是在成功的 rollout 中达到的已验证延续模式的多样性。它依赖于镜像状态、rollout 协议、验证器和模式映射。
  • 实验设置:使用 Qwen2.5-Instruct 系列模型在 GSM8K 上进行推理实验。每个问题采样 32 次 rollout,设置两种反思视界,并对比 96 和 160 两个 Token 预算。
  • 关键发现:随着 Token 预算增加,零可达性降低,验证模式熵增加。Qwen2.5-1.5B 在 160 Token 时实现了最强的镜像视界,证明了在受限反射协议下,较小但结构更优的模型可能具备更强的持续推理能力。

行业启示

  • 评估体系重构:行业应从单一的性能指标转向多维度的动态能力评估,特别关注模型在长程推理和多步反思中的稳定性与多样性。
  • 小模型优化方向:在资源受限场景下,优化模型的“反思协议”和“验证机制”可能比单纯增加参数量更能提升复杂任务的表现,应重视推理过程中的结构稳定性。
  • 算法设计启发:开发新的训练目标或推理策略时,可引入“可行路径熵”等概念,鼓励模型生成更多样且可验证的推理路径,而非仅仅追求单一正确答案。

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。

Research 科学研究 Evaluation 评测 Alignment 对齐