Nemotron DAG-of-Thoughts: A Hybrid Solving Pipeline for Reasoning Competitions

The WeebOrWeed team proposes a LangGraph-based DAG-of-Thoughts architecture, combining deterministic solvers with LLM fallback strategies to efficiently solve six types of puzzles in the NVIDIA reasoning competition (bit operation deduction, cipher decryption, equation transformation, gravity physics calculation, number base conversion, unit conversion). This balances the reasoning ability of LLMs with computational accuracy, improving efficiency and correctness.

The NVIDIA Nemotron Model Reasoning Competition is a challenging AI event on the Kaggle platform, requiring the solution of six types of complex reasoning puzzles characterized by multi-step reasoning and dependence on previous results. Traditional end-to-end LLMs tend to accumulate errors and have weak numerical computation capabilities; the core challenge is to design a hybrid system that balances LLM reasoning and computational accuracy.

The problem is decomposed into sub-nodes using DAG (independent subtasks are executed in parallel). The execution flow has three stages: Classification (identify puzzle type via keyword matching, generate DAG and assign tools), Decomposition (pass through the classifier's DAG for the first time; LLM generates a new DAG upon retry), and Solving (execute ready nodes in parallel, retry if failed).

Five types of puzzles use deterministic solvers: gravity calculation (g=2d/t²), unit conversion (regular expression extraction + factor multiplication), base conversion (Roman numeral lookup table), cipher decryption (character mapping + permutation search), bit operations (bit-level Boolean function search); equation transformation first tries the deterministic mode, and if it fails, LLM multi-round voting is used (take the majority from 7 rounds).

Node-level parallelism is implemented using ThreadPoolExecutor (1-3 threads per round); Retry mechanism: When a node fails, pass the context to LLM to generate a new decomposition strategy (rephrase subproblems, merge steps, etc.), with a maximum of 3 retries; if parsing fails, automatically fall back to a single node.

Deploy Nemotron 3 Nano (4B/30B versions) locally using Ollama without API keys; most puzzles do not call LLMs (instant, free, accurate); engineering value: determinism first, layered fallback, parallelization, failure-driven learning.

This pipeline proves that a hybrid architecture (LLM + deterministic solvers) can build an efficient and reliable system. Project link: https://github.com/WeebOrWeed/Nemotron; Competition link: https://www.kaggle.com/competitions/nvidia-nemotron-model-reasoning-challenge; Tech stack: Python · LangGraph · Ollama · Nemotron 3 Nano · ThreadPoolExecutor.