A transformer architecture alteration to incentivise externalised reasoning

• URL: http://arxiv.org/abs/2603.21376v2
• Date: Tue, 24 Mar 2026 06:21:38 GMT
• Title: A transformer architecture alteration to incentivise externalised reasoning
• Abstract summary: We augment an existing transformer architecture with an early-exit mechanism at intermediate layers and train the model to exit at shallower layers when the next token can be predicted without deep computation.<n>We provide preliminary results to this effect for small reasoning models, showing that they learn to adaptively reduce computations across tokens.<n>We predict that, applied at the right scale, our approach can minimise the amount of excess computation that reasoning models have at their disposal to perform non-myopic planning.
• Score: 1.2571323258597842
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: We propose a new architectural change, and post-training pipeline, for making LLMs more verbose reasoners by teaching a model to truncate forward passes early. We augment an existing transformer architecture with an early-exit mechanism at intermediate layers and train the model to exit at shallower layers when the next token can be predicted without deep computation. After a calibration stage, we incentivise the model to exit as early as possible while maintaining task performance using reinforcement learning. We provide preliminary results to this effect for small reasoning models, showing that they learn to adaptively reduce computations across tokens. We predict that, applied at the right scale, our approach can minimise the amount of excess computation that reasoning models have at their disposal to perform non-myopic planning using their internal activations, reserving this only for difficult-to-predict tokens.

Related papers

• On Neural Scaling Laws for Weather Emulation through Continual Training [76.70310203875866]
  We study neural scaling in Scientific Machine Learning, focusing on models for weather forecasting.<n>To analyze scaling behavior in as simple a setting as possible, we adopt a minimal, scalable, general-purpose Swin Transformer architecture.<n>We show that models trained in this minimalist way follow predictable scaling trends and even outperform standard cosine learning rate schedules.
  arXiv  Detail & Related papers  (2026-03-26T17:37:25Z)
• The Diminishing Returns of Early-Exit Decoding in Modern LLMs [22.442245764201406]
  We re-evaluate layer-wise early-exit in modern Large Language Model (LLM) inference.<n>We introduce a metric to quantify a model's intrinsic suitability for early-exit.<n>We find that dense transformers generally offer greater early-exit potential than Mixture-of-Experts and State Space Models.
  arXiv  Detail & Related papers  (2026-03-24T20:38:17Z)
• Learning a Generative Meta-Model of LLM Activations [75.30161960337892]
  We create "meta-models" that learn the distribution of a network's internal states.<n>Applying the meta-model's learned prior to steering interventions improves fluency, with larger gains as loss decreases.<n>These results suggest generative meta-models offer a scalable path toward interpretability without restrictive structural assumptions.
  arXiv  Detail & Related papers  (2026-02-06T18:59:56Z)
• The Effectiveness of Approximate Regularized Replay for Efficient Supervised Fine-Tuning of Large Language Models [17.1510128169152]
  LoRA-based supervised fine-tuning can catastrophically degrade model capabilities.<n>Small tweaks to the training procedure with very little overhead can virtually eliminate the problem.
  arXiv  Detail & Related papers  (2025-12-26T18:55:42Z)
• Scaling LLM Speculative Decoding: Non-Autoregressive Forecasting in Large-Batch Scenarios [76.85739138203014]
  We present SpecFormer, a novel architecture that accelerates unidirectional and attention mechanisms.<n>We demonstrate that SpecFormer achieves lower training demands and reduced computational costs.
  arXiv  Detail & Related papers  (2025-11-25T14:20:08Z)
• Continuous Autoregressive Language Models [56.49239051750678]
  We introduce Continuous Autoregressive Language Models (CALM)<n>CALM uses a high-fidelity autoencoder to compress a chunk of K tokens into a single continuous vector.<n>We develop a comprehensive likelihood-free framework that enables robust training, evaluation, and controllable sampling.
  arXiv  Detail & Related papers  (2025-10-31T17:58:11Z)
• LESA: Learnable LLM Layer Scaling-Up [57.0510934286449]
  Training Large Language Models (LLMs) from scratch requires immense computational resources, making it prohibitively expensive.<n>Model scaling-up offers a promising solution by leveraging the parameters of smaller models to create larger ones.<n>We propose textbfLESA, a novel learnable method for depth scaling-up.
  arXiv  Detail & Related papers  (2025-02-19T14:58:48Z)
• Training dynamic models using early exits for automatic speech recognition on resource-constrained devices [15.879328412777008]
  Early-exit architectures enable the development of dynamic models capable of adapting their size and architecture to varying levels of computational resources and ASR performance demands. We show that early-exit models trained from scratch not only preserve performance when using fewer encoder layers but also exhibit enhanced task accuracy compared to single-exit or pre-trained models. Results provide insights into the training dynamics of early-exit architectures for ASR models.
  arXiv  Detail & Related papers  (2023-09-18T07:45:16Z)
• Uncovering mesa-optimization algorithms in Transformers [61.06055590704677]
  Some autoregressive models can learn as an input sequence is processed, without undergoing any parameter changes, and without being explicitly trained to do so. We show that standard next-token prediction error minimization gives rise to a subsidiary learning algorithm that adjusts the model as new inputs are revealed. Our findings explain in-context learning as a product of autoregressive loss minimization and inform the design of new optimization-based Transformer layers.
  arXiv  Detail & Related papers  (2023-09-11T22:42:50Z)
• Dynamic Context Pruning for Efficient and Interpretable Autoregressive Transformers [29.319666323947708]
  We present a novel approach that dynamically prunes contextual information while preserving the model's expressiveness. Our method employs a learnable mechanism that determines which uninformative tokens can be dropped from the context. Our reference implementation achieves up to $2times$ increase in inference throughput and even greater memory savings.
  arXiv  Detail & Related papers  (2023-05-25T07:39:41Z)
• Confident Adaptive Language Modeling [95.45272377648773]
  CALM is a framework for dynamically allocating different amounts of compute per input and generation timestep. We demonstrate the efficacy of our framework in reducing compute -- potential speedup of up to $times 3$ -- while provably maintaining high performance.
  arXiv  Detail & Related papers  (2022-07-14T17:00:19Z)

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.