Research map / Large language models / Reasoning and chain of thought

Deep Reasoning in General Purpose Agents via Structured Meta-Cognition

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces Deep Reasoning, a formal language for structured meta-reasoning, and the Dolores agent that builds task-specific recursive scaffolds from in-context atomic decomposition examples, outperforming ReAct, CodeAct, Deep Research and RLM baselines on four long-horizon reasoning benchmarks.

Problem

LLM agent scaffolds hard-code how reasoning is decomposed, making them brittle when the task requires a different reasoning structure. The paper targets flexible, task-adaptive decomposition of complex reasoning for general-purpose agents.

Approach

A formal language (Deep Reasoning) that represents meta-reasoning as executable decompositions mixing associative (LLM) calls, formal (Python) computation, and recursive sub-agent calls; human meta-reasoning traces are formalized into atomic decompositions and supplied as in-context examples so the modeling LLM constructs the scaffold at test time.

Key results

• Dolores with Qwen3-32B reports 0.346 on SynthWorlds, 0.369 on PhantomWiki, and 0.132 on Oolong-real, outperforming the best baseline in those settings.
• Averaged over settings, the paper reports Dolores improving on the strongest baseline by 24.8%, with an 8B Dolores surpassing 32B baselines in more than half of settings.
• Ablation shows removing in-context decomposition examples causes about 70% average drop, and natural-language principles without examples performs worse than no examples.

Surprising findings

• Replacing the in-context decomposition examples with natural-language decomposition principles performs even worse than having no examples at all.
• CodeAct outperforms Dolores on SynthWorlds with Llama-3.3-70B, the one exception among 12 settings.

Limitations

• Writing good Deep Reasoning decompositions requires expertise in the language, LLM abilities and agent loops, a barrier to entry for domain experts.
• High token cost of spawning many reasoning threads.
• Complex scientific or medical tasks may require many atomic decompositions that could overwhelm a single memory thread.

Experimental setup

Datasets: SynthWorlds PhantomWiki DeepSearchQA Oolong-real

Benchmarks: SynthWorlds PhantomWiki DeepSearchQA Oolong-real

Metrics: F1 accuracy relaxed accuracy

Baselines: ReAct CodeAct Deep Research RLM

Builds on: ReAct CodeAct Recursive Language Models

Closest papers by meaning

• A$^2$FM: An Adaptive Agent Foundation Model for Tool-Aware Hybrid Reasoning (ICLR 2026 · similarity 0.89)
• Reasoning Scaffolding: Distilling the Flow of Thought from LLMs (ICLR 2026 · similarity 0.88)
• Eliciting Reasoning in Language Models with Cognitive Tools (NeurIPS 2025 · similarity 0.88)
• Trade-offs in Large Reasoning Models: An Empirical Analysis of Deliberative and Adaptive Reasoning over Foundational Capabilities (AAAI 2026 · similarity 0.88)
• LYNX: Learning Dynamic Exits for Confidence-Controlled Reasoning (NeurIPS 2026 · similarity 0.88)
• DeepCompress: A Dual Reward Strategy for Dynamically Exploring and Compressing Reasoning Chains (ICLR 2026 · similarity 0.88)