Chain-of-Thought Provably Enables Learning the (Otherwise) Unlearnable

Authors: Chenxiao Yang, Zhiyuan Li, David Wipf

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we demonstrate our proposed Co T construction significantly enhances the reasoning capabilities of real-world LLMs in solving challenging arithmetic reasoning tasks, including learning polynomials and Boolean formulas. ... 5 EXPERIMENTS
Researcher Affiliation Collaboration Chenxiao Yang , Zhiyuan Li , David Wipf Toyota Technological Institute at Chicago, Amazon Web Services EMAIL, EMAIL
Pseudocode Yes Algorithm 1: Step-by-Step Learning with Co T (ACo T)
Open Source Code Yes Codes are available at https://github.com/chr26195/Co T-ICL.
Open Datasets No For empirical verification, we consider new arithmetic reasoning tasks and test if the decomposition schemes from our analysis are practically effective (Section 5). Specifically, we propose to construct complex reasoning tasks with varying overall hardness and hardness of subtasks. ... For each k in parities and w in DNFs, we similarly i.i.d. sample 100 target functions.
Dataset Splits Yes For each target function, the LLMs are provided with 10 demonstrations and asked to infer the computation process and apply it to derive the output for an unseen input. ... For each function, we provide LLMs with 100 in-context examples, ask them to find patterns in these examples, and return the output for each query.
Hardware Specification No The paper mentions evaluating
Software Dependencies No The paper mentions evaluating
Experiment Setup Yes For each target function, the LLMs are provided with 10 demonstrations and asked to infer the computation process and apply it to derive the output for an unseen input. ... We test 100 times to compute the success rate. ... For the input values, we uniformly select two unique integers from the set {2, 3, . . . , 10}. ... For each k in parities and w in DNFs, we similarly i.i.d. sample 100 target functions. ... we implement it by randomly masking H 1 consecutive intermediate steps.