Graph-Native Reinforcement Learning Enables Traceable Scientific Hypothesis Generation through Conceptual Recombination

Subhadeep Pal; Shashwat Sourav; Tirthankar Ghosal; Markus J. Buehler

Source-linked AI summary

Graph-Native Reinforcement Learning Enables Traceable Scientific Hypothesis Generation through Conceptual Recombination

Subhadeep Pal, Shashwat Sourav, Tirthankar Ghosal, Markus J. Buehler

arXiv:2607.00924v1Jul 1, 2026cs.AIcond-mat.mtrl-scics.CLcs.LG

TL;DR

Materials-science hypothesis generation requires connecting fragmented concepts, mechanisms, and evidence in ways that linear language-model reasoning does not explicitly represent. Graph-PRefLexOR uses graph-native structured reasoning and outperforms corresponding base models on 100 open-ended questions, with stronger traceability and reasoning–answer alignment.

Problem
Materials-science hypothesis generation requires organizing fragmented entities, mechanisms, constraints, and outcomes, while standard language-model traces lack explicit relational and causal structure.
Method
Graph-PRefLexOR uses GRPO-trained reasoning stages for exploration, graph construction, pattern extraction, and hypothesis synthesis, linking language generation with inspectable symbolic structure.
Results
Graph-PRefLexOR consistently outperformed corresponding base models across 100 open-ended materials-science and mechanics questions, with strongest gains in reasoning traceability and stronger reasoning–answer alignment.
Takeaways & Limitations
Graph-native reasoning provides an interpretable framework for scientific hypothesis generation with traceable intermediate pathways and iterative conceptual recombination.
Takeaways & Limitations
The reward design imposes competing structural terms, so their absolute levels are not directly comparable and total reward saturates below 1.0 by construction.

Abstract

from arXiv · show

Accelerating materials discovery requires AI systems that can generate scientifically valid hypotheses through multi-step, domain-grounded reasoning. Standard large language models often produce fluent but weakly traceable responses to open-ended materials design problems, making it difficult to determine whether final answers are supported by coherent intermediate reasoning. We develop Graph-PRefLexOR, a family of graph-native reasoning models fine-tuned with Group Relative Policy Optimization (GRPO) to organize reasoning into explicit phases for mechanism exploration, graph construction, pattern extraction, and hypothesis synthesis. This design links neural language generation with symbolic relational structure, enabling causal connections to be constructed, inspected, and reused. On 100 open-ended questions from materials science and mechanics literature, Graph-PRefLexOR achieves 40-65% improvements over corresponding base models, with the largest gains in reasoning traceability. Embedding analyses show broader semantic exploration and approximately 2-3 times greater semantic diversity than baselines. Semantic backtracking and layer-wise hidden-state analyses further show stronger alignment between structured reasoning and final answers. Finally, test-time graph expansion reveals that additional compute primarily increases long-range conceptual recombination within a bounded semantic space, rather than simply expanding semantic coverage. These results establish graph-native reinforcement learning as a pathway toward interpretable AI systems for scientific hypothesis generation in materials design and other scientific applications.

1 Introduction

Materials-science hypothesis generation requires connecting mechanisms and evidence across scales and domains, while standard LLM reasoning is difficult to trace. Graph-PRefLexOR addresses this gap by exposing structured reasoning phases that reorganize how hypotheses are generated and recombined.

Motivation: Scientific hypothesis generation in materials science and mechanics requires linking molecular structure, mesoscale organization, interfaces, defects, processing history, and boundary conditions.Macroscopic properties emerge from coupled processes spanning these scales and factors.
Motivation: Standard LLM responses can be difficult to trace, increasing risks of untraceability, hallucination, or contradiction in scientific reasoning.This limitation motivates explicit intermediate representations for mechanism and relation construction.
Related work: Graph representations make scientific relationships inspectable by representing concepts as nodes and their relationships through explicit mechanisms, dependencies, and analogies.Retrieval-augmented generation, knowledge graphs, and agentic workflows address complementary parts of the broader problem.
Method: Graph-PRefLexOR exposes a sentinel-based trace with <brainstorm>, <graph>, <graph_json>, <patterns>, and <synthesis> phases for structured scientific reasoning.The phases separate mechanism exploration, concept abstraction, and graph construction before synthesis.
Evaluation: The evaluation uses 100 manually curated open-ended questions from materials science and mechanics literature covering cross-domain linkage, causal mapping, hidden variables, model abstraction, and hypothesis generation.The benchmark targets capabilities that standard factual or multiple-choice evaluations do not adequately assess.
Contribution: The approach changes hypothesis generation by restructuring the intermediate computational pathway through which hypotheses are generated, aligned, and iteratively recombined.The contribution concerns intermediate reasoning organization rather than only changing final answers.

2 Results and Discussion

Graph-PRefLexOR consistently improves aggregate performance and reasoning traceability over corresponding base models, while explicit phase-based graph reasoning produces more structured, diverse, and directionally organized latent trajectories. Despite these different reasoning paths, the models converge toward semantically similar final answers.

Benchmark performance: 40–65% aggregate improvements across model scales over corresponding baselines, with the largest gains in Reasoning Traceability.Figure 2 compares three graph-native GRPO variants with their corresponding base models across three evaluation metrics.
Reasoning dependence: 30–50% overall performance reductions when reasoning is disabled, indicating that explicit reasoning drives most observed gains beyond architectural differences.The degradation closely mirrors that of the Llama baseline in the no-thinking setting.
Scale effects: The 8B model scores approximately 25–30% higher than the 1.7B variant across all metrics, consistent with greater capacity for expressive graph construction and pattern extraction.The comparison is reported in Fig. 2d across model scales.
Traceable reasoning: Phase-separated reasoning converts mechanism exploration into entities, relations, patterns, and synthesis, producing an inspectable graph that links immune-system concepts to multi-agent AI mechanisms.The trace and extracted representations are shown in Figures 4 and 5, including the proposed bridge from adaptive memory expansion to clonal selection and long-term robustness.
Latent-space dynamics: Graph-PRefLexOR reasoning trajectories are broader, more directional, and phase-separated, whereas baselines remain localized and entangled; final-answer embeddings nevertheless show tight clustering and strong overlap.The structured model diversifies earlier before synthesis, while both model classes converge toward semantically similar endpoints.
Semantic organization: 2.9× and 2.6× gains in mean inter-phase cosine distance at 1.7B and 8B, respectively, show greater semantic diversity than corresponding base models.Distances increase from 0.07 to 0.20 at 1.7B and from 0.08 to 0.21 at 8B, indicating stronger differentiation among reasoning stages.

3 Conclusion

Graph-PRefLexOR uses GRPO and staged graph-native reasoning to generate traceable scientific hypotheses. Across 100 materials-science and mechanics questions, it improves reasoning quality and traceability while anchoring answers to structured traces and enabling iterative conceptual recombination.

3 Conclusion: Graph-PRefLexOR combines GRPO with sentinel-based <brainstorm>, <graph>, <graph_json>, <patterns>, and <synthesis> stages for interpretable scientific hypothesis generation.These stages decompose reasoning into mechanism exploration, concept abstraction, machine-readable graph construction, pattern extraction, and final hypothesis synthesis.
3 Conclusion: Across 100 open-ended materials-science and mechanics questions, Graph-PRefLexOR outperformed corresponding base models in reasoning quality, intellectual depth, and reasoning traceability.The strongest improvements were observed in traceability, indicating that graph-structured reasoning enhances the organization and causal transparency of intermediate reasoning.
3 Conclusion: Graph-PRefLexOR final answers remain strongly anchored to their own structured reasoning pathway, most frequently backtracking to the <synthesis> stage.By contrast, Qwen3-8B answers align with their own visible thinking traces in only a minority of cases and more often approach Graph-PRefLexOR-derived outputs semantically.
3 Conclusion: Accumulated <graph_json> outputs and expansion policies turn Graph-PRefLexOR into a self-expanding graph engine for iterative test-time ideation.Additional test-time compute does not simply expand semantic territory indefinitely, but instead changes the explored embedding volume and maximum explored distance.
3 Conclusion: Null-model analyses show enriched relation-typed motifs, beyond-chance modularity, and systematic links between semantically dissimilar concepts, with novelty concentrated in two-hop conceptual bridges.Direct graph edges remain mildly homophilic, whereas two-hop bridges connect normally unrelated concepts.

4 Materials and Methods … Stage 1: ORPO Cold Start

Graph-PRefLexOR uses a two-stage, graph-native training recipe across 1.7B, 3B, and 8B models, beginning with ORPO cold-start alignment from structured preference pairs. Its reasoning format proceeds from exploration to graph formalization, pattern abstraction, and synthesis, making intermediate reasoning parseable and inspectable.

4.1 Training Strategy: Graph-PRefLexOR trains 1.7B, 3B, and 8B models with one two-stage recipe applied to reasoning and standard instruction-tuned backbones.Qwen3 backbones adapt native reasoning abilities, whereas Llama-3.2-3B-Instruct receives the graph-native format.
4.1.1 Dataset Construction: Teacher distillation converts streaming general and domain-specific text into structured graph-native preference records containing questions, answers, chosen traces, rejected responses, and extracted graphs.Malformed graph_json or empty-answer examples are discarded, ensuring retained records contain valid graphs and complete answers.
Stage 1: ORPO Cold Start: Figure 18 shows ORPO loss declining and preference accuracy saturating near 1.0 across all three backbones, with the 1.7B model’s larger reward margin attributed to its higher learning rate.The ORPO runs last approximately 480, 480, and 240 steps for 1.7B, 3B, and 8B, respectively.
4.1.1 Dataset Construction: The preference construction contrasts a rich graph-reasoning trace with a shallow direct answer, making the ordering easy to learn during ORPO cold start.Graph-GRPO later uses only the prompt and gold answer, leaving chosen and rejected fields unused.
4.1.2 Training Approach: The model emits a structured trace ordered as <brainstorm>, <graph>, <graph_json>, <patterns>, and <synthesis>, followed by the final answer outside </think>.The sequence encodes exploration → formalization → abstraction → explanation rather than free-form chain-of-thought.
4.1.2 Training Approach: Graph-GRPO training uses composite rewards over the structured traces, with graph utility remaining the lowest reward component at every model scale.Figure 19 reports that 8B starts highest while 3B climbs most across differing training durations.
Rationale for This Reasoning Structure: Explicit graph structure supports relational faithfulness, verifiability, and answer faithfulness because graph_json is parseable, reward-accessible, and must support the final synthesis.Validity, structure, diversity, and graph-utility terms operate directly on the canonical graph object.
Stage 1: ORPO Cold Start: ORPO aligns each backbone using a preferred-response NLL term and an odds-ratio penalty against the dispreferred response, without a frozen reference model.The stage uses one epoch, a 5% held-out split, and seed 42; it establishes reliable format adherence before reinforcement learning.

Stage 2: Graph-GRPO · Semantic rewards (judge-graded) · Format reward

Stage 2 applies critic-free GRPO with composite rewards to train graph-native reasoning while keeping updates memory-efficient through group sampling and LoRA adaptation. The reward combines judge-graded answer quality and graph utility with programmatic graph and format checks, including parsing gates that cap malformed outputs.

Stage 2: Graph-GRPO: GRPO samples G = 8 completions per prompt, computes group-normalized advantages from composite rewards, and optimizes a clipped policy objective with a KL penalty.The method uses no learned value network and generates through vLLM for throughput.
Stage 2: Graph-GRPO: LoRA-only adaptation keeps GRPO updates lightweight and mitigates catastrophic forgetting when optimizing the black-box reward.The reward combines judge calls and graph analytics rather than providing a differentiable signal.
Semantic rewards (judge-graded): Each completion receives a scalar reward formed as a fixed convex combination of six components normalized to [0, 1].The trainer’s headline total reward is the per-step mean across completions.
Semantic rewards (judge-graded): Two reward components use an external LLM judge, while four are computed programmatically from the parsed graph G = (V, E).The external judge is grok-4-1-fast-non-reasoning.
Semantic rewards (judge-graded): Correctness directly grades the post-think answer against the gold answer, while graph utility grades an answer reconstructed solely from the emitted graph_json.Graph utility therefore tests whether the graph contains sufficient information to reproduce the answer without outside knowledge.
Format reward: Format reward credits the required reasoning sections, parseable graph_json, patterns, synthesis, and a non-empty node set.Section credits are think 0.15, brainstorm 0.10, graph 0.15, graph_json 0.20, patterns 0.15, synthesis 0.15, and non-empty nodes 0.10.
Format reward: Malformed graph_json gates graph_json and subsequent credits, capping the format score when parsing fails.The parsing requirement makes structural validity a prerequisite for later format components.

NetworkX-validity reward · Diversity reward · Structure reward

The reward combines graph validity, semantic diversity, and topology to favor internally consistent, non-collapsed, connected, hierarchical reasoning graphs. Its structure terms explicitly shape graph size, density, internal inferences, reasoning depth, and connectivity.

NetworkX-validity reward: The validity reward is defined using invalid-edge and self-loop quantities to assess graph consistency.Einv denotes edges referencing nonexistent nodes, while ℓ denotes the number of self-loops.
NetworkX-validity reward: Eval = m −|Einv| rewards internally consistent, connected graphs by penalizing edges that reference nonexistent nodes.The validity formulation uses Einv for invalid edge references; the surrounding definition also tracks self-loops through ℓ.
Diversity reward: Diversity is measured from mean off-diagonal cosine similarity among Sentence-BERT embeddings of textual graph elements.The embedded elements include node ids and source-relation-target triples, using all-MiniLM-L6-v2.
Diversity reward: A richness bonus b = min(0.1, m′/100) penalizes degenerate graphs of near-duplicate nodes that could reward-hack validity and structure.The bonus is capped at 0.1 and is intended to discourage collapsed graph representations.
Structure reward: The topology reward rstruct clips a sum of size, density, internal-node, depth, and connectivity terms to [0,1].The size term peaks for graphs with 5–20 nodes, while density is based on directed density ρ.
Structure reward: The structure reward uses directed density sdens = min(0.2, 2ρ) and an internal-node term sint = 0.3 nint.These terms reward graph topology containing directed relations and nodes that mediate inferences.
Structure reward: Depth and weak-connectivity terms shape graphs into connected, hierarchical scaffolds, with longest DAG path L representing reasoning-chain length.Internal nodes correspond to intermediate inferences, while L captures the length of the reasoning chain.

Design rationale … Optimization Diagnostics

The reward design prioritizes semantic correctness and graph utility while using programmatic shaping terms to enforce valid, diverse, structured traces. Across model sizes, ORPO installs the reasoning format before Graph-GRPO improves rewards, with optimization behavior shaped by backbone reasoning ability and reward dispersion.

Design rationale: 0.55 of the reward weights correctness and graph utility, while 0.45 weights programmatic validity, diversity, and structure terms for dense shaping and anti-hacking.Format and NetworkX-validity can reach 1.0, whereas diversity and structure terms are soft-capped; total reward therefore saturates below 1.0 by construction.
Training Dynamics and Results: Across all three sizes, cold start establishes the reasoning format and preference ordering before Graph-GRPO improves composite reward, with backbone type determining headroom.The backbone and model size govern starting conditions and the amount of reward improvement available.
Graph-PRefLexOR-1.7B: ∼2.11→∼1.38 ORPO loss, 0.95→1.0 preference accuracy, and 0.14→∼1.0 reward margin characterize Qwen3-1.7B training.The NLL nearly coincides with total loss because the odds-ratio penalty is negligible under the larger 5 × 10−5 learning rate; the large margin reflects learning rate rather than model scale.
Graph-PRefLexOR-3B: ∼2.1→∼1.47 ORPO loss, ∼0.02–0.03 component separation, ≈1.0 preference accuracy, and 0.03→0.16 reward margin characterize Llama-3.2-3B-Instruct.The instruction backbone converges cleanly within one epoch, plateauing after ∼300 steps, with the same accuracy on the held-out split.
Graph-PRefLexOR-8B: ∼1.69→∼1.31 ORPO loss in ∼240 steps and ≈1.0 preference accuracy from the outset characterize Qwen3-8B, whose reward rises modestly from ∼0.60 to ∼0.63.Its strong base separates preferences immediately, so cold start mainly installs output format; the reward also dips mid-run to ∼0.60.
Generation-length dynamics: The 1.7B and 8B models lengthen completed traces to ∼2.6k and ∼2.4k tokens with ≤3% truncation, whereas the 3B reduces truncation from ∼20% to ∼1%.The 3B’s reinforcement-learning dynamics simultaneously shorten its mean termination behavior, while the other models rarely exhaust the 8000-token budget.
Optimization Diagnostics: ∼0.11 initial reward dispersion for 3B versus ∼0.05–0.06 for 1.7B and 8B gives 3B the strongest GRPO learning signal, which tapers toward ∼0.07.Policy entropy decreases modestly for all models, zero-variance groups remain at 0, and the gradient therefore stays usable throughout.

4.2 Benchmark Question Generation · 4.3 Answer Backtracking and Hidden-State Analysis

Sections 4.2–4.3 construct a 100-question open-ended scientific reasoning benchmark and define reproducible semantic-backtracking and hidden-state analyses to trace how model reasoning relates to final answers.

4.2 Benchmark Question Generation: The benchmark pipeline converts papers to Markdown, extracts high-level mechanistic fields, and generates questions from domains including materials science and large language models.Marker performs layout-aware conversion, while gpt-4o-mini extracts titles, DOIs, abstracts, results, discussions, and conclusions; introductions, methods, and references are excluded.
4.2 Benchmark Question Generation: The benchmark contains 100 self-contained, research-level questions spanning five predefined scientific reasoning categories.Questions are generated from research papers and refined without changing their systems, variables, causal structures, question types, or intended reasoning challenges.
4.2 Benchmark Question Generation: A second gpt-5.4 refinement pass improves readability, grammar, precision, and benchmark suitability while preserving each question’s scientific reasoning challenge.The editing step avoids introducing new scientific claims or simplifying the task and outputs a polished JSONL benchmark.
4.3 Answer Backtracking and Hidden-State Analysis: Semantic backtracking assigns each final answer to the candidate reference with the highest cosine similarity in normalized BGE embedding space.The method compares embedded final answers with candidate texts and uses deterministic candidate ordering to break numerical ties rather than applying an absolute similarity threshold.
4.3 Answer Backtracking and Hidden-State Analysis: For Qwen3-8B, backtracking compares final answers with the model’s thinking trace and four Graph-PRefLexOR stages, reporting full and binary source distributions.The binary split distinguishes backtracking to Qwen’s own thinking trace from backtracking to another source.
4.3 Answer Backtracking and Hidden-State Analysis: For Graph-PRefLexOR 8B, cross-model backtracking compares its final answer with four internal stages and Qwen3-8B outputs, while internal-only analysis isolates the closest structured stage.The internal stages are <brainstorm>, <graph>, <patterns>, and <synthesis>; the cross-model analysis also reports whether answers backtrack internally or to Qwen outputs.
4.3 Answer Backtracking and Hidden-State Analysis: Layer-wise hidden-state analysis measures reasoning–answer distances for Qwen3-8B and Graph-PRefLexOR 8B, including stage-specific comparisons and means with one-standard-deviation bands.Additional linear-probe and logit-lens-style analyses test whether reasoning–answer distinctions are linearly decodable and compare token preferences across spans; these analyses are observational.

4.4 Iterative Graph-native Ideation and Scaling Analysis · Expansion Strategies

The model is evaluated as a self-expanding ideation engine in which structured reasoning traces become graph updates that generate later questions. Four expansion policies allocate test-time compute toward graph frontiers, semantic peripheries, distant recombinations, or language-level follow-ups.

4.4 Iterative Graph-native Ideation and Scaling Analysis: Each iteration answers the current question, parses its reasoning trace into typed nodes and labeled relations, and merges the local graph into a global directed graph.This creates an iterative loop in which generated reasoning both expands the graph and determines subsequent ideation.
Expansion Strategies: The four runs differ only in how each expansion policy maps the accumulated graph to the next batch of questions, allocating compute to distinct idea-space frontiers.The policies use graph degree, betweenness, node embeddings, or language-level questioning to direct expansion.
Expansion Strategies: The frontier strategy selects low-degree unvisited nodes plus the highest-betweenness hub, then asks follow-ups about unresolved mechanisms.It combines outward exploration of less-developed regions with consolidation around a structurally central concept.
Expansion Strategies: The novelty strategy explores the embedding periphery by selecting nodes least aligned with the current centroid for follow-up questioning.This directs computation toward semantically peripheral concepts rather than graph-central structure.
Expansion Strategies: The leap strategy connects peripheral concepts to their most dissimilar graph partners and imports unrelated-field principles to force mechanistic recombination and cross-domain transfer.These operations encourage concepts and connections that are distant from the existing graph.
Expansion Strategies: The converse strategy uses a separate questioner to propose unanchored follow-ups from the seed question and latest answer, introducing absent concepts beyond saturated regions at two calls per iteration.Its prompts target implications, tensions, cross-domain analogies, or deeper mechanisms, unlike the node- or node-pair-anchored alternatives.

Scaling of Surprising-insight Yield · Growth Dynamics

The analyses quantify how surprising conceptual recombination accumulates with test-time compute and characterize graph growth through embedding geometry and community structure. They distinguish long-range recombination from homophilic links and separate conceptual novelty from consolidating in-fill.

Scaling of Surprising-insight Yield: The scaling analysis reconstructs accumulated graphs at 40 compute checkpoints from t = 0 to 2000 without rerunning the model.Nodes and edges are retained when their birth iteration is ≤ t, with node labels embedded once using google/embeddinggemma_300m.
Scaling of Surprising-insight Yield: It measures graph-size-robust quantities including distinct concepts to quantify how insight yield scales with test-time compute.The supplied passage introduces four robust metrics, but only the distinct-concept metric is visible here.
Scaling of Surprising-insight Yield: Surprising recombinations are defined using a global null distribution based on the exact mean µ and standard deviation σ of final-graph pairwise cosine similarities.Concept pairs are considered atypical when their combination score is less than −1.
Scaling of Surprising-insight Yield: The cumulative yield counts atypical concept pairs bridged through a shared intermediate concept at graph distance two, excluding directly linked pairs.Each bridged pair receives the first iteration at which it is realized, determined by the maximum birth iteration of its two endpoints.
Growth Dynamics: New concepts are classified as novel or consolidating in-fill using their embedding distance from the running concept centroid, with exploration radius measured from the seed.Arrivals are binned by iteration, and each bin reports the novel-concept fraction plus mean and interquartile range of seed distances.
Growth Dynamics: Edge replay across 30 checkpoints tracks greedy-modularity communities, modularity Q, and recombination edges after the graph becomes connected.A recombination edge links endpoints already connected by a prior path, while community counts and Q describe mesoscale sub-field formation and interconnection.

Statements and Declarations

The work was primarily supported by the U.S. Department of Energy’s SciDAC FORUM-AI project, with no declared competing interests. Training data, benchmark data, code, models, and analysis artifacts are available through linked repositories or upon request.

Funding: Primary support came from the U.S. Department of Energy’s SciDAC program under the FORUM-AI project.Support came through the Office of Science, including Advanced Scientific Computing Research and Basic Energy Sciences.
Competing interests: The authors declare no competing financial or non-financial interests relevant to the article.
Data and code: Training data, benchmark data, analysis artifacts, and full training and analysis code are publicly available through Hugging Face and GitHub.Additional materials are also available from the corresponding author upon reasonable request.
Model availability: The trained Graph-PRefLexOR 8B, 3B, and 1.7B models are available through Hugging Face.The model repositories are provided for Graph-Preflexor-8b_12292025, Graph-Preflexor-3b_08012026, and Graph-Preflexor-1.7b_08012026.
Use of LLMs: Large language models supported benchmark construction, response generation, question refinement, and independent evaluation, with all generated materials reviewed, filtered, and analyzed by the authors.

SUPPLEMENTARY INFORMATION

The supplementary information identifies the authors and their affiliations across MIT, Oak Ridge National Laboratory, Washington University in St. Louis, and Lawrence Berkeley National Laboratory.

Subhadeep Pal and Markus J. Buehler are affiliated with MIT’s Department of Civil and Environmental Engineering.
Tirthankar Ghosal is affiliated with Oak Ridge National Laboratory’s Computing and Computational Sciences Directorate.
Shashwat Sourav is affiliated with Washington University in St. Louis and Oak Ridge National Laboratory’s Computing and Computational Sciences Directorate.
The listed affiliations also include Lawrence Berkeley National Laboratory in Oak Ridge, Tennessee, USA.
Markus J. Buehler is also affiliated with MIT’s Department of Mechanical Engineering and Schwarzman College of Computing.

S1 Representative Qwen3-8B Baseline Response

This section presents the full Qwen3-8B response to a representative benchmark question, illustrating the baseline’s linear, extended, and repetitive reasoning style. It provides a comparison point for Graph-PRefLexOR-8B’s structured, phase-wise reasoning.

The full Qwen3-8B response serves as a representative baseline for comparison with Graph-PRefLexOR-8B’s phase-wise reasoning.The section uses the response to contrast linear baseline reasoning with structured reasoning.
The baseline begins by comparing biological immune systems with multi-agent AI frameworks through a sequential recall of their components and functions.It discusses B-cells, T-cells, memory cells, signaling, adaptation, and feedback loops.

S2 Qwen3-8B Behavior Analysis on Layers 7–10

In Qwen3-8B, layers 7–10 mark a representational transition where thinking and answer states become more geometrically separated, while later layers carry more directly recoverable answer content. These analyses support interpreting graph-structured reasoning as improving the stability and interpretability of reasoning-to-answer generation.

Qwen3-8B layer-wise behavior: Layers 7–10 form an early transition zone where thinking and answer representations become more geometrically separated, without storing the final answer itself.The model has entered answer-generation mode, but answer content is not yet fully consolidated.
Qwen3-8B layer-wise behavior: Near-perfect probe AUROC precedes layers 7–10, whereas hidden-state divergence rises there and spikes at the final layer.Probe decodability indicates that generation mode is identifiable early, while divergence measures later geometric separation.
Qwen3-8B layer-wise behavior: Later layers, particularly around layer 30, recover final-answer similarity more strongly than layers 7–10 through activation patching.This suggests that later representations contain more directly recoverable answer-specific content.
Qwen3-8B layer-wise behavior: The analyses clarify that intermediate layers diagnose the transition from visible reasoning to answer formation, while later layers carry explicit answer content.Probes and representation distances reveal separability or presence in hidden states but do not by themselves establish causal use.
Qwen3-8B layer-wise behavior: Graph-structured reasoning is positioned as improving final-answer quality and the stability and interpretability of the pathway from reasoning to response generation.This conclusion follows from the contrast between reasoning traces and final responses emphasized in the analysis.