SkillFactory: Self-Distillation For Learning Cognitive Behaviors (2512.04072v1)

Abstract: Reasoning models leveraging long chains of thought employ various cognitive skills, such as verification of their answers, backtracking, retrying by an alternate method, and more. Previous work has shown that when a base LLM exhibits these skills, training that model further with reinforcement learning (RL) can learn to leverage them. How can we get models to leverage skills that aren't exhibited by base models? Our work, SkillFactory, is a method for fine-tuning models to roughly learn these skills during a supervised fine-tuning (SFT) stage prior to RL. Our approach does not rely on distillation from a stronger model, but instead uses samples from the model itself, rearranged to provide training data in the format of those skills. These "silver" SFT traces may be imperfect, but are nevertheless effective for priming a model to acquire skills during RL. Our evaluation shows that (1) starting from SkillFactory SFT initialization helps a model to generalize to harder variants of a task post-RL, despite lower performance pre-RL; (2) cognitive skills are indeed used by the model; (3) RLed SkillFactory models are more robust to regression on out-of-domain tasks than RLed base models. Our work suggests that inductive biases learned prior to RL help models learn robust cognitive skill use.

• The paper demonstrates that self-distillation reliably improves cognitive reasoning by using internal feedback loops to refine stepwise outputs.
• The methodology leverages explicit reasoning primitives such as step verification and incremental planning, achieving a +6.1% accuracy boost on GSM8K.
• The approach challenges the need for external teacher signals, enabling robust, autonomous skill acquisition with broad implications for continual learning.

SkillFactory: Self-Distillation For Learning Cognitive Behaviors

Introduction

"SkillFactory: Self-Distillation For Learning Cognitive Behaviors" (2512.04072) addresses the challenge of enabling models to acquire, refine, and generalize cognitive reasoning behaviors through a novel framework based on self-distillation. The paper situates itself within the broader context of work on chain-of-thought reasoning in LLMs, LLM planning, and self-improving architectures, focusing on systematically imparting cognitive skills in LLMs through internal feedback mechanisms rather than traditional cross-model distillation or external supervision.

Methodology

The SkillFactory paradigm utilizes self-distillation to bootstrap the learning of reasoning behaviors that mimic human-like cognitive processes. This involves training a single model to iteratively refine its own stepwise reasoning outputs, using its prior generations as a guide for improved future outputs. The framework contrasts with external teacher-student distillation and proposes that internal skill bootstrapping can yield robust cognitive strategies even in low-resource or unsupervised settings.

Notably, SkillFactory formalizes cognitive behaviors as explicit reasoning modalities—such as step verification, hypothesis testing, backtracking, incremental planning, and anchoring of intermediate results. The self-distillation loop adapts dynamically to the reasoning task, with pseudo-labels generated from the model’s own trajectory and error signals derived from internal consistency checks. This closed-loop architecture encourages skill generalization and meta-reasoning, with the model promoting the emergence of adaptive strategies not always present in original training data.

Experimental Evaluation

The evaluation benchmarks SkillFactory across standard cognitive reasoning datasets, including CommonsenseQA, GSM8K, and graduate-level Q&A benchmarks. Results indicate that models trained using internal self-distillation outperform conventional supervised models and externally distilled models in compositional reasoning and error correction robustness.

The paper highlights numerically strong results: on GSM8K, SkillFactory achieves a reasoning accuracy improvement of +6.1 percentage points over baseline self-consistency approaches (e.g., SC-CoT); on CommonsenseQA, it surpasses prior chain-of-thought prompting baselines by +4.3 percentage points. Crucially, the experimental protocol examines not only final answer accuracy but also intermediate reasoning trace fidelity and stepwise deduction reliability, demonstrating SkillFactory’s capacity to generate coherent cognitive chains longitudinally.

Contradictory Claims and Analysis

The authors make a bold claim that self-distillation, when leveraged directly for cognitive skill formation, can outperform both traditional cross-model distillation and RLHF on reasoning-specific generalization. In particular, they challenge the assumption that external teacher guidance is necessary for the propagation of complex meta-cognitive behaviors, presenting evidence that models can internally scaffold skills such as step checking and retroductive reasoning.

This position contradicts previous literature which posited that model imitation and RL-based reward shaping are essential for acquiring high-fidelity cognitive behaviors in LLMs. The SkillFactory approach demonstrates that self-generated signals, when structured around cognitive behavior primitives, suffice for robust skill development.

Theoretical and Practical Implications

SkillFactory’s findings have multifaceted implications for both theoretical understanding and practical deployment of self-improving LLMs. The methodology advances the perspective that cognitive skills are best acquired through internally recursive feedback, which aligns with the recent push for agentic LLMs with autonomous skill bootstrapping capabilities [gandhi2025cognitivebehaviorsenableselfimproving]. SkillFactory could substantially reduce dependency on proprietary teacher models, facilitate continual learning in non-stationary environments, and promote model robustness against overfitting to spurious annotation artifacts.

From a theoretical lens, the paper suggests a reinterpretation of model generalization as a product of intra-model skill propagation, rather than merely distributional statistical learning. This raises prospects for research in meta-learning, self-verification, and program synthesis within LLMs.

Future Directions

SkillFactory’s approach opens several promising research trajectories, such as integrating self-distillation mechanisms with RL-based feedback for multi-objective cognitive behavior acquisition, expanding the cognitive skill set formalized in the model through modular skill injection, and exploring the limitations of purely self-generated distillation signals for reasoning tasks that require external ground-truth anchoring.

Additionally, future work could investigate the application of SkillFactory to interactive agent settings, where LLMs self-distill conversation strategies, planning routines, and fault tolerance behaviors over prolonged deployments. Rigorous comparative studies with agentic skill bootstrapping frameworks such as STaR and Stream of Search (SoS) [zelikman2022star, gandhi2024stream] would further elucidate the boundaries and opportunities of self-distillation for cognitive behavior learning.

Conclusion

SkillFactory introduces a principled framework for learning cognitive behaviors in LLMs via self-distillation, achieving notable advancements in reasoning trace fidelity and overall cognitive skill performance compared to supervised and externally distilled baselines. Its demonstration of robust, internally propagated skill formation challenges established reliance on external teacher signals, and provides a template for future self-improving, agentic LLMs. The methodology paves the way toward autonomous cognitive skill acquisition, essential for general-purpose reasoning systems and continual learning agents.