Towards Understanding Distilled Reasoning Models: A Representational Approach

• URL: http://arxiv.org/abs/2503.03730v2
• Date: Tue, 25 Mar 2025 00:07:50 GMT
• Title: Towards Understanding Distilled Reasoning Models: A Representational Approach
• Authors: David D. Baek, Max Tegmark,
• Abstract summary: We train a crosscoder on Qwen-series models and their fine-tuned variants.<n>Our results suggest that the crosscoder learns features corresponding to various types of reasoning, including self-reflection and verification.
• Score: 6.563993791037387
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: In this paper, we investigate how model distillation impacts the development of reasoning features in large language models (LLMs). To explore this, we train a crosscoder on Qwen-series models and their fine-tuned variants. Our results suggest that the crosscoder learns features corresponding to various types of reasoning, including self-reflection and computation verification. Moreover, we observe that distilled models contain unique reasoning feature directions, which could be used to steer the model into over-thinking or incisive-thinking mode. In particular, we perform analysis on four specific reasoning categories: (a) self-reflection, (b) deductive reasoning, (c) alternative reasoning, and (d) contrastive reasoning. Finally, we examine the changes in feature geometry resulting from the distillation process and find indications that larger distilled models may develop more structured representations, which correlate with enhanced distillation performance. By providing insights into how distillation modifies the model, our study contributes to enhancing the transparency and reliability of AI systems.

Related papers

• Style over Substance: Distilled Language Models Reason Via Stylistic Replication [4.313454680394974]
  Specialized reasoning language models (RLMs) have demonstrated that scaling test-time computation through detailed reasoning traces significantly enhances performance. In this study, we investigate to what extent distilled models internalize replicated stylistic patterns during reasoning. We find that models trained on the synthetic traces achieve comparable performance, indicating that distilled reasoning abilities rely significantly on surface-level patterns.
  arXiv  Detail & Related papers  (2025-04-02T13:50:20Z)
• On the Reasoning Capacity of AI Models and How to Quantify It [0.0]
  Large Language Models (LLMs) have intensified the debate surrounding the fundamental nature of their reasoning capabilities.<n>While achieving high performance on benchmarks such as GPQA and MMLU, these models exhibit limitations in more complex reasoning tasks.<n>We propose a novel phenomenological approach that goes beyond traditional accuracy metrics to probe the underlying mechanisms of model behavior.
  arXiv  Detail & Related papers  (2025-01-23T16:58:18Z)
• The Buffer Mechanism for Multi-Step Information Reasoning in Language Models [52.77133661679439]
  Investigating internal reasoning mechanisms of large language models can help us design better model architectures and training strategies. In this study, we constructed a symbolic dataset to investigate the mechanisms by which Transformer models employ vertical thinking strategy. We proposed a random matrix-based algorithm to enhance the model's reasoning ability, resulting in a 75% reduction in the training time required for the GPT-2 model.
  arXiv  Detail & Related papers  (2024-05-24T07:41:26Z)
• Diffusion Model with Cross Attention as an Inductive Bias for Disentanglement [58.9768112704998]
  Disentangled representation learning strives to extract the intrinsic factors within observed data. We introduce a new perspective and framework, demonstrating that diffusion models with cross-attention can serve as a powerful inductive bias. This is the first work to reveal the potent disentanglement capability of diffusion models with cross-attention, requiring no complex designs.
  arXiv  Detail & Related papers  (2024-02-15T05:07:54Z)
• Revealing Multimodal Contrastive Representation Learning through Latent Partial Causal Models [85.67870425656368]
  We introduce a unified causal model specifically designed for multimodal data. We show that multimodal contrastive representation learning excels at identifying latent coupled variables. Experiments demonstrate the robustness of our findings, even when the assumptions are violated.
  arXiv  Detail & Related papers  (2024-02-09T07:18:06Z)
• Leveraging Model-based Trees as Interpretable Surrogate Models for Model Distillation [3.5437916561263694]
  Surrogate models play a crucial role in retrospectively interpreting complex and powerful black box machine learning models. This paper focuses on using model-based trees as surrogate models which partition the feature space into interpretable regions via decision rules. Four model-based tree algorithms, namely SLIM, GUIDE, MOB, and CTree, are compared regarding their ability to generate such surrogate models.
  arXiv  Detail & Related papers  (2023-10-04T19:06:52Z)
• Knowledge Distillation Performs Partial Variance Reduction [93.6365393721122]
  Knowledge distillation is a popular approach for enhancing the performance of ''student'' models. The underlying mechanics behind knowledge distillation (KD) are still not fully understood. We show that KD can be interpreted as a novel type of variance reduction mechanism.
  arXiv  Detail & Related papers  (2023-05-27T21:25:55Z)
• GAM(e) changer or not? An evaluation of interpretable machine learning models based on additive model constraints [5.783415024516947]
  This paper investigates a series of intrinsically interpretable machine learning models. We evaluate the prediction qualities of five GAMs as compared to six traditional ML models.
  arXiv  Detail & Related papers  (2022-04-19T20:37:31Z)
• Beyond Explaining: Opportunities and Challenges of XAI-Based Model Improvement [75.00655434905417]
  Explainable Artificial Intelligence (XAI) is an emerging research field bringing transparency to highly complex machine learning (ML) models. This paper offers a comprehensive overview over techniques that apply XAI practically for improving various properties of ML models. We show empirically through experiments on toy and realistic settings how explanations can help improve properties such as model generalization ability or reasoning.
  arXiv  Detail & Related papers  (2022-03-15T15:44:28Z)
• Beyond Trivial Counterfactual Explanations with Diverse Valuable Explanations [64.85696493596821]
  In computer vision applications, generative counterfactual methods indicate how to perturb a model's input to change its prediction. We propose a counterfactual method that learns a perturbation in a disentangled latent space that is constrained using a diversity-enforcing loss. Our model improves the success rate of producing high-quality valuable explanations when compared to previous state-of-the-art methods.
  arXiv  Detail & Related papers  (2021-03-18T12:57:34Z)
• Improving the Reconstruction of Disentangled Representation Learners via Multi-Stage Modeling [54.94763543386523]
  Current autoencoder-based disentangled representation learning methods achieve disentanglement by penalizing the ( aggregate) posterior to encourage statistical independence of the latent factors. We present a novel multi-stage modeling approach where the disentangled factors are first learned using a penalty-based disentangled representation learning method. Then, the low-quality reconstruction is improved with another deep generative model that is trained to model the missing correlated latent variables.
  arXiv  Detail & Related papers  (2020-10-25T18:51:15Z)
• Learning Invariances for Interpretability using Supervised VAE [0.0]
  We learn model invariances as a means of interpreting a model. We propose a supervised form of variational auto-encoders (VAEs) We show how combining our model with feature attribution methods it is possible to reach a more fine-grained understanding about the decision process of the model.
  arXiv  Detail & Related papers  (2020-07-15T10:14:16Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.