Transcoder Adapters for Reasoning-Model Diffing

• URL: http://arxiv.org/abs/2602.20904v1
• Date: Tue, 24 Feb 2026 13:40:28 GMT
• Title: Transcoder Adapters for Reasoning-Model Diffing
• Abstract summary: We introduce transcoder computation, a technique for learning an interpretable approximation.<n>We apply transcoder adapters to characterize the differences between Qwen2.5-Math-7B and its reasoning-distilled variant, DeepSeek-R1-Di-Qwen-7B.
• Score: 28.090312772463253
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: While reasoning models are increasingly ubiquitous, the effects of reasoning training on a model's internal mechanisms remain poorly understood. In this work, we introduce transcoder adapters, a technique for learning an interpretable approximation of the difference in MLP computation before and after fine-tuning. We apply transcoder adapters to characterize the differences between Qwen2.5-Math-7B and its reasoning-distilled variant, DeepSeek-R1-Distill-Qwen-7B. Learned adapters are faithful to the target model's internal computation and next-token predictions. When evaluated on reasoning benchmarks, adapters match the reasoning model's response lengths and typically recover 50-90% of the accuracy gains from reasoning fine-tuning. Adapter features are sparsely activating and interpretable. When examining adapter features, we find that only ~8% have activating examples directly related to reasoning behaviors. We deeply study one such behavior -- the production of hesitation tokens (e.g., "wait"). Using attribution graphs, we trace hesitation to only ~2.4% of adapter features (5.6k total) performing one of two functions. These features are necessary and sufficient for producing hesitation tokens; removing them reduces response length, often without affecting accuracy. Overall, our results provide insight into reasoning training and suggest transcoder adapters may be useful for studying fine-tuning more broadly.

Related papers

• To Adapt or Not to Adapt? Selective Adaptation for Vision-Language Models [50.94061118407214]
  Test-time adaptation (TTA) has emerged as a prominent strategy for adapting vision-language models to distribution shifts during inference.<n>We conduct a per-sample analysis of model predictions before and after adaptation, and observe two failure modes in existing TTA methods.<n>We introduce a new problem of selective adaptation, which aims to determine whether a given test sample should undergo adaptation or be skipped.
  arXiv  Detail & Related papers  (2026-09-08T07:35:45Z)
• Correcting Suppressed Log-Probabilities in Language Models with Post-Transformer Adapters [0.0]
  Post-transformer adapter corrects suppression on 31 ideology-discriminating facts across Qwen3-4B, 8B, and 14B.<n> adapter memorizes all 15 training facts and generalizes to 11--39% of 16 held-out facts across 5 random splits per scale.
  arXiv  Detail & Related papers  (2026-03-25T16:27:43Z)
• Learning Self-Interpretation from Interpretability Artifacts: Training Lightweight Adapters on Vector-Label Pairs [1.028001761831691]
  Self-interpretation methods prompt language models to describe their own internal states.<n>We show that training lightweight adapters on interpretability artifacts, while keeping the LM entirely frozen, yields reliable self-interpretation.
  arXiv  Detail & Related papers  (2026-02-10T22:50:02Z)
• Correct, Concise and Complete: Multi-stage Training For Adaptive Reasoning [11.179446105672461]
  We propose a multi-stage efficient reasoning method that combines supervised fine-tuning and reinforcement learning.<n>Our approach reduces response length by an average of 28% for 8B models and 40% for 32B models.<n>It achieves a superior trade-off compared to more complex state-of-the-art efficient reasoning methods.
  arXiv  Detail & Related papers  (2026-01-06T12:31:51Z)
• TokenSqueeze: Performance-Preserving Compression for Reasoning LLMs [57.217593337454026]
  TokenSqueeze is a novel Long2Short method that condenses reasoning paths while preserving performance and relying exclusively on self-generated data.<n>We show that TokenSqueeze reduces token usage while maintaining accuracy on the MATH500 benchmark.
  arXiv  Detail & Related papers  (2025-11-17T10:38:56Z)
• Reasoning Vectors: Transferring Chain-of-Thought Capabilities via Task Arithmetic [51.41777906371754]
  Large language models often require costly optimization, such as reinforcement learning, to master complex reasoning tasks.<n>This work demonstrates that reasoning ability, once learned, can be extracted and transferred between models as a compact task vector.
  arXiv  Detail & Related papers  (2025-09-01T11:04:51Z)
• SEAL: Steerable Reasoning Calibration of Large Language Models for Free [58.931194824519935]
  Large Language Models (LLMs) have demonstrated compelling capabilities for complex reasoning tasks via the extended chain-of-thought (CoT) reasoning mechanism.<n>Recent studies reveal substantial redundancy in the CoT reasoning traces, which negatively impacts model performance.<n>We introduce SEAL, a training-free approach that seamlessly calibrates the CoT process, improving accuracy while demonstrating significant efficiency gains.
  arXiv  Detail & Related papers  (2025-04-07T02:42:07Z)
• Unveiling Reasoning Thresholds in Language Models: Scaling, Fine-Tuning, and Interpretability through Attention Maps [3.8936716676293917]
  This study investigates the in-context learning capabilities of various decoder-only transformer-based language models with different model sizes and training data.<n>We identify a critical parameter threshold (1.6 billion), beyond which reasoning performance improves significantly in tasks such as commonsense reasoning in multiple-choice question answering and deductive reasoning.
  arXiv  Detail & Related papers  (2025-02-21T00:48:32Z)
• Parameter-Efficient Sparse Retrievers and Rerankers using Adapters [4.9545244468634655]
  We study adapters for SPLADE, a sparse retriever, for which adapters retain the efficiency and effectiveness otherwise achieved by finetuning. We also address domain adaptation of neural retrieval thanks to adapters on cross-domain BEIR datasets and TripClick.
  arXiv  Detail & Related papers  (2023-03-23T12:34:30Z)
• LegoNet: A Fast and Exact Unlearning Architecture [59.49058450583149]
  Machine unlearning aims to erase the impact of specific training samples upon deleted requests from a trained model. We present a novel network, namely textitLegoNet, which adopts the framework of fixed encoder + multiple adapters'' We show that LegoNet accomplishes fast and exact unlearning while maintaining acceptable performance, synthetically outperforming unlearning baselines.
  arXiv  Detail & Related papers  (2022-10-28T09:53:05Z)
• VisFIS: Visual Feature Importance Supervision with Right-for-the-Right-Reason Objectives [84.48039784446166]
  We show that model FI supervision can meaningfully improve VQA model accuracy as well as performance on several Right-for-the-Right-Reason metrics. Our best performing method, Visual Feature Importance Supervision (VisFIS), outperforms strong baselines on benchmark VQA datasets. Predictions are more accurate when explanations are plausible and faithful, and not when they are plausible but not faithful.
  arXiv  Detail & Related papers  (2022-06-22T17:02:01Z)
• PRover: Proof Generation for Interpretable Reasoning over Rules [81.40404921232192]
  We propose a transformer-based model that answers binary questions over rule-bases and generates the corresponding proofs. Our model learns to predict nodes and edges corresponding to proof graphs in an efficient constrained training paradigm. We conduct experiments on synthetic, hand-authored, and human-paraphrased rule-bases to show promising results for QA and proof generation.
  arXiv  Detail & Related papers  (2020-10-06T15:47:53Z)
• Tailoring: encoding inductive biases by optimizing unsupervised objectives at prediction time [34.03150701567508]
  Adding auxiliary losses to the main objective function is a general way of encoding biases that can help networks learn better representations. In this work we take inspiration from textittransductive learning and note that after receiving an input, we can fine-tune our networks on any unsupervised loss. We formulate em meta-tailoring, a nested optimization similar to that in meta-learning, and train our models to perform well on the task objective after adapting them using an unsupervised loss.
  arXiv  Detail & Related papers  (2020-09-22T15:26:24Z)

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.