fluid-reasoning-representation-phase1
收藏资源简介:
该数据集是Fluid Reasoning Representation(Hook等人)论文用于ARLR 2026反驳的Phase 1成果,旨在支持机制可解释性、推理和规划领域的研究,特别是探索流体推理表示。它扩展了原始的QwQ x Mystery Blocksworld研究,主要扩展内容包括:1) 引入了第二个大型推理模型:Llama-3.3-Nemotron-Super-49B-v1;2) 新增了两个研究领域:Mystery Logistics(使用混淆动作/谓词词汇的PDDL Logistics领域)和GSM8K-Renamed(对表面名词和动词进行混淆处理的数学文字问题领域)。数据集由16个单元格构成,每个单元格对应一个特定的(模型,领域)组合,并包含相应的跟踪数据(Traces),数据大小从0.7 MB到9.4 MB不等,且均已进行表示提取(Reps标记为yes)。数据生成过程使用vLLM 0.20.2,采用贪婪解码(温度=0),每个单元格基于200个提示生成。表示提取使用Hugging Face Transformers,在模型默认的引导层(QwQ-32B为第47层,Nemotron-49B为第60层)设置前向钩子获取激活。跟踪数据在分词前被截断至前20-40个段落以控制内存。本阶段新增了C3因果引导过程:对于每个混淆跟踪,在引导层安装钩子,将混淆短语的激活替换为对应的干净命名中心激活(位置配对),并比较在三种条件下(基线无干预、替换为干净中心、消融为零)跟踪自身后续令牌的负对数似然(NLL),以验证流体推理表示特征(即干净中心比零值干扰更小)。数据集适用于文本生成、机制可解释性、推理和规划相关的研究任务,采用Apache-2.0许可证。
This dataset is the Phase 1 achievement from the paper *Fluid Reasoning Representation* by Hook et al., intended for the ARLR 2026 rebuttal. It aims to support research in mechanistic interpretability, reasoning and planning, particularly the exploration of fluid reasoning representations. It extends the original QwQ x Mystery Blocksworld study, with key extensions as follows: 1) Introduction of a second large reasoning model: Llama-3.3-Nemotron-Super-49B-v1; 2) Addition of two new research domains: Mystery Logistics (the PDDL Logistics domain with obfuscated action and predicate vocabulary) and GSM8K-Renamed (the math word problem domain with obfuscated surface nouns and verbs). The dataset consists of 16 cells, each corresponding to a specific (model, domain) pair and containing corresponding trace data (Traces). The size of each cell ranges from 0.7 MB to 9.4 MB, and all have undergone representation extraction (marked with Reps=yes). The data generation process uses vLLM 0.20.2 with greedy decoding (temperature=0), and each cell is generated based on 200 prompts. For representation extraction, Hugging Face Transformers is used, with forward hooks installed at the model's default guidance layers (layer 47 for QwQ-32B, layer 60 for Nemotron-49B) to capture activations. Trace data is truncated to the first 20-40 paragraphs before tokenization to control memory usage. This phase adds the C3 causal guidance process: for each obfuscated trace, hooks are installed at the guidance layers to replace the activations of obfuscated phrases with their corresponding position-aligned clean-named centroid activations. The negative log likelihood (NLL) of the subsequent tokens of the trace itself is compared under three conditions: baseline with no intervention, replacement with clean centroids, and setting to zero, to verify the fluid reasoning representation characteristics (i.e., clean centroids introduce less interference than zero values). The dataset is suitable for research tasks related to text generation, mechanistic interpretability, reasoning and planning, and is released under the Apache-2.0 license.




