Deep Reasoning in General Purpose Agents via Structured Meta-Cognition

The gist Deep Reasoning introduces an inference-time approach for constructing task-specific scaffolds through structured meta-reasoning, which allows general-purpose agents to adapt their reasoning structure just-in-time based on the task requirements.

Deep Reasoning Language

The paper introduces a formal language called Deep Reasoning that represents meta-reasoning as executable decompositions over associative inference, formal computation, and recursive subproblem solving (Page 3). This language distinguishes between three dimensions of reasoning: Associative vs. Formal (D1), Object vs. Meta-level (D2), and Atomic vs. Monolithic decomposition of meta-reasoning (D3) (Page 3). The language defines symbols, an interpretation function f, and specific interpreters like the Associative Interpreter A and the Formal Interpreter E, which allows for explicit distinction between these reasoning modes (Page 5).

DOLORES Agent Architecture

The Deep Reasoning agent instantiated is DOLORES (Deep Layered Reasoning Scaffold), which recursively defines its operation through the equation ID(x) = EID, A(x) (Page 4). The modeling function m, implemented as an LLM, is responsible for decomposing tasks into a series of associative steps (calls to A), formal steps (code executed by E), and meta-reasoning steps (recursive calls to ID) (Page 7). DOLORES uses in-context meta-reasoning examples derived from human traces to guide its reasoning and dynamically adapt its scaffold at test time, which leads to a nested cascade of formal, associative and meta-reasoning steps (Page 7).

Operationalizing Meta-Reasoning

The agent operationalizes human meta-reasoning by grounding it in human meta-reasoning traces which are mapped into the language of Deep Reasoning (Page 7). These formalized traces are converted directly into in-context examples that guide the modeling LLM m (Page 7). The system uses these examples to define how the task should be decomposed, enabling fine-grained control over decomposition patterns without additional programming (Page 23).

Performance and Failure Analysis

DOLORES outperforms state-of-the-art scaffold methods across four hard benchmarks by an average of 24.8% over the best-performing baseline (Page 1). The agent succeeds by distributing cognition across structured, lower-load reasoning threads, which reduces premature termination and hallucination (Page 1). Failure analysis shows that dominant failure modes for baselines are premature termination (78%) and hallucination (45%) because they delegate too much to a single LLM context thread (Page 8). DOLORES avoids these pitfalls by decomposing tasks into atomic associative, formal, or meta-reasoning steps that can be handled reliably by a single LLM memory thread (Page 1).

Ablation and Impact

Ablation studies on Qwen3-32B show that removing the in-context examples causes large drops on every benchmark (70% on average), suggesting current LLMs are not good at operationalizing structured meta-reasoning and require humans to meta-reason about task decompositions (Page 22). DOLORES's impact has the potential to lower the barrier for domain experts by encoding expertise as a handful of in-context decompositions rather than as training data or predefined scaffolds (Page 14).

The paper demonstrates that LLMs should not be viewed as a collection of shallow experts, but rather as associative reasoning processes that flexibly interact with formal processes to create artificial minds that can meta-reason robustly (Page 7). The results point toward future agentic systems that treat scaffolding as adaptive reasoning, constructing the structure each task requires just-in-time (Page 1).

Deep Reasoning in General Purpose Agents via Structured Meta-Cognition

Humans solve complex problems by flexibly shifting among reasoning modes, often without explicit deliberation: they plan, execute, revise intermediate goals, resolve ambiguity through associative judgment, and apply formal procedures to well-specified subproblems.

We introduce Deep Reasoning – an inference-time approach for constructing task-specific scaffolds through structured meta-reasoning.

Deep Reasoning uses a formal language that represents meta-reasoning as executable decompositions over associative inference, formal computation, and recursive subproblem solving, enabling decomposition principles to be encoded as in-context examples that guide test-time scaffold construction.

We instantiate this approach in a general-purpose agent (DOLORES) that distributes complex tasks across smaller, more controlled reasoning threads while preserving dependencies among subproblems.

DOLORES outperforms all evaluated scaffolds across four hard benchmarks, three model sizes, and two model families, improving over the strongest evaluated scaffold baseline by 24.8% on average, including methods tailored to individual benchmark families.

Trace and token analyses suggest that while baseline scaffolds fail by overloading individual LLM calls, DOLORES succeeds by distributing cognition across structured, lower-load reasoning threads, thereby reducing premature termination and hallucination.

This advantage can even bridge the scaling gap, with an 8B version surpassing all evaluated 32B baselines from the same family in more than half the settings.

These results point toward future agentic systems that treat scaffolding as adaptive reasoning, constructing the structure each task requires just-in-time.

Prior work characterizes reasoning along two orthogonal dimensions: how reasoning is carried out and what the reasoning is about.

Associative Reasoning generally operates through intuitive proximity shaped by memory and context [Mednick, 1962, Sloman, 1996].

Formal Reasoning proceeds via the sequential application of logical rules over structured, welldefined representations [Wason, 1968, Rips, 1983, JohnsonLaird and Byrne, 1991].

Humans tightly interleave meta- and object-level reasoning when solving tasks that go beyond their cognitive capacity (total processing resources available in working memory).

Our strong reasoning and generalization abilities are often attributed to these meta-reasoning and modeling capabilities [Kargupta et al., 2025].

In order to instill strong reasoning and meta-reasoning capacities in agentic systems, we need to view reasoning processes in terms of the cognitive capacity and abilities of LLMs.

We need to take our intuitive reasoning ability and precisely formalize it and separate it along three dimensions: D1: Associative vs. Formal – delegating what we can to formal code.

D2: Object vs. Meta – decomposing object-level tasks using meta-reasoning until each associative object-level task fits within the capacity of an LLMs.

D3: Atomic vs. Monolithic – decomposing the meta-reasoning, which is itself cognitively demanding, into composable atomic units that can be delegated to different context threads.

In the language of Deep Reasoning, the distinction between associative and formal reasoning is explicit.

When humans meta-reason about a task, we create a step-by-step decomposition on how to solve it (Figure 1).

We can express delegation between formal and associative tasks (D1, §3.1) and even formalize dependencies between associative sub tasks using A and EA.</ref:2605.

1. Deep Reasoning language allows for decomposing meta-reasoning into atomic units using associative reasoning, formal computation, and recursive subproblem solving. This enables agents to construct task-specific scaffolds just-in-time rather than relying on fixed inference patterns that are brittle when tasks require adapting the structure of reasoning itself.

2. The DOLORES agent can distribute complex tasks across smaller, more controlled reasoning threads while preserving dependencies among subproblems, which reduces premature termination and hallucination because it structures reasoning in more fine-grained ways.

3. The system can bridge scaling gaps by using cheaper models for associative tasks and selecting models based on the cognitive load of those tasks, as demonstrated by the 8B model outperforming all evaluated 32B baselines in more than half of settings.

4. The agent can handle long-context aggregation and deep research-style information seeking effectively by decomposing these tasks into atomic steps that are manageable within a single LLM context thread, mitigating the issue of overloading individual LLM calls.

5. Agents can perform complex multi-hop reasoning over synthetic knowledge graphs or long real-world documents by recursively chunking them inside a programmatic environment via recursive self-calls, as seen in the RLM agent's approach to long inputs.

Humans intuitively solve complex problems by flexibly shifting among reasoning modes: they plan, execute, revise intermediate goals, resolve ambiguity through associative judgment, and apply formal procedures to well-specified subproblems. Current LLM agents lack this flexibility, as their scaffolds hard-code such reasoning decisions in advance. These scaffolds are effective when their prescribed structure matches the task, but brittle when solving the task requires adapting the structure of reasoning itself. We introduce Deep Reasoning -- an inference-time approach for constructing task-specific scaffolds through structured meta-reasoning. Deep Reasoning uses a formal language that represents meta-reasoning as executable decompositions over associative inference, formal computation, and recursive subproblem solving, enabling decomposition principles to be encoded as in-context examples that guide test-time scaffold construction. We instantiate this approach in a general-purpose agent (DOLORES) that distributes complex tasks across more controlled reasoning threads. We evaluate it against state-of-the-art scaffolding methods across four hard benchmarks: multi-hop reasoning, long-chain question answering, long-context aggregation, and deep research-style information seeking. DOLORES outperforms all evaluated scaffolds across three model sizes and two model families, improving over the strongest evaluated scaffold baseline by 24.8% on average. DOLORES distributes cognition across structured, lower-load reasoning threads, thereby reducing premature termination and hallucinations. This advantage can even bridge the scaling gap, with an 8B version surpassing all evaluated 32B baselines from the same family in more than half the settings. These results point toward future agentic systems that treat scaffolding as adaptive reasoning, constructing the structure each task requires just-in-time.