CogFlow Framework for Cognitive Reasoning

• CogFlow is a cognitive-inspired framework that decomposes complex reasoning into discrete, interpretable units for visual and social tasks.
• It applies hierarchical stages and multi-objective reward formulations to ensure coherence, fidelity, and structured decision-making.
• Empirical results demonstrate improved accuracy, efficiency, and decision quality over conventional models in both visual and social domains.

CogFlow is a cognitive-inspired framework for structured reasoning, instantiated in two principal research tracks: (1) visual mathematical problem solving via hierarchical perception-internalization-reasoning with explicit knowledge grounding (Chen et al., 5 Jan 2026), and (2) social decision-making via a tree-structured composition of cognitive units to interpret ambiguous social scenarios (Zhou et al., 26 Sep 2025). In both settings, CogFlow defines discrete, interpretable stages or units of cognition and enforces fidelity across stages using multi-objective reward formulations within supervised and reinforcement learning pipelines. Empirical results demonstrate that CogFlow yields increased coherence, efficiency, and decision quality in domains that require complex cognitive integration.

1. Foundational Principles and Motivating Problems

CogFlow is introduced to address the limitations of LLMs in domains where shallow, direct inference is insufficient and where interpretability, stepwise reasoning, and fidelity to observed or extracted facts are non-negotiable. In visual mathematical problem solving, extant models often fail to ground reasoning in actual visual input, leading to "reasoning drift" where the answer sequence loses connection to perceived entities, such as geometric primitives or diagram semantics (Chen et al., 5 Jan 2026). In social cognition, conventional stepwise deduction misses the inherently interpretive and multi-faceted nature of human social reasoning, requiring explicit modeling of observation, attribution, motivation, emotion regulation, perceived efficacy, and intended behavior (Zhou et al., 26 Sep 2025).

CogFlow posits that high-quality reasoning in these settings must explicitly (a) modularize cognitive processes, (b) maintain traceability between perception and inference, and (c) provide reinforcement signals that reward not just correctness but also faithfulness and diversity in the reasoning chain.

2. Structural Composition: Stages and Units

Visual Mathematical Reasoning

CogFlow decomposes problem solving into three sequential stages:

• Perception (“Watching”): Input is the raw diagram $I$; output is a set of normalized visual primitives $y^\pi_{\rm per}$ (e.g., points, lines, circles). This stage isolates geometric cues needed for subsequent processing.
• Knowledge Internalization (“Thinking”): This stage transforms $y^\pi_{\rm per}$ into a formal, symbolic representation $y^\pi_{\rm int}$, explicitly binding named entities (e.g., “AB is a diameter”).
• Reasoning (“Answering”): The reasoning module receives $y^\pi_{\rm int}$ and the original question, outputs a multi-step chain-of-thought $\{s_t\}$ and final answer $y^\pi_{\rm ans}$.

Each stage is connected via explicit reward signals that encourage not only correct outcomes but also faithfulness to perceptual inputs and avoidance of visually ungrounded reasoning (Chen et al., 5 Jan 2026).

Social Reasoning

CogFlow in the social domain formalizes cognitive flows consisting of six units, grounded in social-cognitive theory (Bandura et al., 1986):

• Observation (O): Interpreting raw social cues and events.
• Attribution (A): Inferring causal relationships or motives.
• Motivation (M): Evaluating agent goals and value functions.
• Regulation (R): Adjusting intermediate beliefs or affective states.
• Efficacy (E): Assessing the likelihood of successful action.
• Behavior (B): Formulating actionable social responses.

A cognitive flow $T$ is then a sequence $T = [(u_1, c_1), (u_2, c_2), ..., (u_n, c_n)]$, where each $u_i$ is a unit type and $y^\pi_{\rm per}$0 is content (Zhou et al., 26 Sep 2025).

3. Reward Design and Learning Objectives

CogFlow instances implement composite rewards that go beyond standard correctness:

Visual Mathematical Reasoning

• Synergistic Visual Reward (SynVR): Composed of
  • Visual Parameterized Reward (VPR): Penalizes geometric deviation between predicted ($y^\pi_{\rm per}$1) and ground truth ($y^\pi_{\rm per}$2) primitives using Hungarian matching over parameter spaces.
  • Visual Semantic Reward (VSR): Rewards semantic style/layout similarity by computing cosine similarity between CLIP embeddings of predicted and true diagrams.
  • The combination is: $y^\pi_{\rm per}$3, with $y^\pi_{\rm per}$4.
• Knowledge Internalization Reward (IntlzR): A contrastive model $y^\pi_{\rm per}$5 penalizes reasoning chains that omit or distort perceptual facts.
• These rewards are injected into a Visual-Gated Policy Optimization (VGPO) routine that regulates the policy to align multi-stage outputs and suppress shortcut reasoning (Chen et al., 5 Jan 2026).

Social Reasoning

• Format Reward ($y^\pi_{\rm per}$6): Enforces production of well-structured tagged flow sequences.
• Comparative Preference Reward ($y^\pi_{\rm per}$7): Uses a learned classifier to prefer responses over references based on multi-turn curation.
• Cognitive Diversity Reward ($y^\pi_{\rm per}$8): Penalizes frequent repetition of cognitive units, incentivizing coverage of all six units when warranted.
• Reasoning Length Reward ($y^\pi_{\rm per}$9): Applies a soft window on reasoning steps to avoid both superficial and overlong chains.
• The total RL reward is: $y^\pi_{\rm per}$0 with default $y^\pi_{\rm per}$1 (Zhou et al., 26 Sep 2025).

4. Data Curation, Architecture, and Optimization Procedures

Visual Mathematical Problem Solving

• Dataset: MathCog, consisting of 120K+ samples with fine-grained perception-reasoning annotated alignments.
• Architecture: Multimodal transformers (e.g., LLMs augmented with perception modules, CLIP embeddings for semantic visual reward)
• Policy Optimization: Visual-Gated Policy Optimization integrates stage-specific rewards.
• Hyperparameters: Stage-specific, with explicit regularization for each module.

Social Reasoning

• Data Source: 5,100 Reddit situations distilled to (S,Q) pairs, tree-structured simulation of flows using the Deepseek-R1 LLM, filtered via preference ranking (CPRank₂) and theory-derived criteria.
• Architecture: Fine-tuned Llama-3.1-8B-Instruct and Qwen-2.5-7B-Instruct models; SFT and RL conducted via LLaMA-Factory and veRL.
• Special Tokens: Six unit tags incorporated into the tokenizer for explicit sequence demarcation.
• Supervised and RL Phases: Initial warm-up with negative log-likelihood minimization, then RL with Group Relative Policy Optimization and multi-objective reward (Zhou et al., 26 Sep 2025).

5. Empirical Evaluation and Key Results

Visual Mathematical Reasoning

• Experiments on visual math benchmarks show that CogFlow achieves higher faithfulness in reasoning chains and improved benchmark scores relative to state-of-the-art multimodal LLMs.
• Ablations demonstrate the utility of both SynVR and Knowledge Internalization rewards, with performance dropping when either is removed (Chen et al., 5 Jan 2026).

Social Reasoning

• Automated Evaluation: On a 500-instance test set, CogFlow achieves a normalized preference score of 0.7828—exceeding DeepSeek-R1 (0.6578) and distilled R1 chains (0.7310).
• Reasoning Efficiency: CogFlow chains are more concise (average 391 tokens vs. 621).
• Human Studies: Cognitive flows are rated higher in coherence (4.44/5), efficiency (4.62/5), interpretability, and predictability.
• Intervention Outcomes: Augmenting human participants with CogFlow hints yields an accuracy improvement of +9.5%, compared to +7.5% for conventional R1-style chains. Grouped intervention shows an +8% gain for those guided by cognitive flows vs. a decline for the control (Zhou et al., 26 Sep 2025).

6. Significance and Implications

CogFlow operationalizes modular, interpretable cognitive flows, providing a reproducible mechanism for bridging perception and reasoning (in vision tasks) and simulating human-like associative processing (in social tasks). The introduction of contrastive and diversity-promoting rewards mitigates reasoner drift and encourages breadth in cognitive modeling. A plausible implication is the generalizability of such architectures to other domains where rigorous grounding and stepwise interpretation are required. Empirical validation confirms improvements not only in machine-selected metrics but also in measures of human-augmented performance and decision quality.