Frontier-to-Open Neuro-Symbolic Transfer: Dissecting and Transferring Constraint Structures in Model-Provided Reasoning Traces
Recent empirical studies of Chain-of-Thought (CoT) reasoning rely heavily on small open-weight Supervised Fine-Tuning (SFT) distilled models (e.g., 7B–8B parameters), often treating their observable CoT behavior as representative of reasoning models at large. However, it remains unclear whether frontier reinforcement-learning (RL) foundation models expose qualitatively different, transferable symbolic structures in their model-provided reasoning traces compared to open-weight distilled models. In this paper, we investigate Frontier-to-Open Neuro-Symbolic Transfer across 1,000 reasoning traces and coupled counterfactual parameter-perturbation tasks spanning three formal constraint classes: algebraic inequalities (\(C_1\)), temporal state-transition automata (\(C_2\)), and multi-source physical conservation laws (\(C_3\)). First, using a two-layer deterministic and Gold-IR grounded audit, we measure the constraint progress curve \(R(k)\) as a function of normalized trace length \(k\). We find that frontier reasoning traces (Claude Opus 5.5 and Gemini 4 Argon) complete 83.9% and 70.9% of pure Gold-IR constraint grounding within the first \(k = 20%\) of their traces, achieving 64.0% and 62.0% counterfactual Pair-Consistent Accuracy, whereas open-weight distilled models (DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-Llama-8B) complete only 13.8% and 11.9% by \(k = 20%\) and collapse to 0.0% and 2.0% Pair-Consistent Accuracy. Second, to disentangle pure constraint grounding from teacher decision hints and formatting effects, we introduce a two-tier symbolic Intermediate Representation: Group D0 (NeSy-Constraint-IR), which encodes only variables, derived quantities, and formal inequality polarity with 0% target action strings, and Group D (NeSy-Decision-IR), which adds explicit rule-deduction scaffolding. Across an eight-group controlled ablation on three open-weight backbones (2,400 GPU evaluations, including unstructured prose Group G and non-JSON structured natural language Group H), we uncover a striking backbone dependency: on general instruction-tuned Qwen2.5-7B-Instruct, pure constraint-only Group D0 reaches 50.0% Pair-Consistent Accuracy (outperforming Group G 26.0% by +24.0 pp, Group H 32.0% by +18.0 pp, and even Group D 46.0%). On SFT-distilled reasoning models (R1-Distill-Qwen-7B and R1-Distill-Llama-8B), Group D0 lifts consistency from 0.0% \(\rightarrow\) 16.0% and 2.0% \(\rightarrow\) 20.0%, while decision-augmented Group D reaches 32.0% and 36.0% (reducing memorization traps to 2.0%–4.0%), demonstrating that SFT-distilled CoT models require explicit rule-deduction scaffolding to overcome post-constraint action-mapping inertia.
Authors
- Wenjian Liu (ORCID: https://orcid.org/0000-0002-9794-5160)
Publication Details
- Journal
- Preprints.org
- Published
- 2026-11-13
- DOI
- https://doi.org/10.20944/preprints202611.1447.v1
- Primary Topic
- Topic Modeling
- Type
- preprint