mlsanityData to 14 Oct 2026

HALO: Hybrid Adaptive Latent Reasoning for Language Models

arXiv 2607.08775 ↗First posted 3 May 2026cs.CL cs.LGcited by 0

On the map: LLM reasoning → Test-time compute and efficient thinking

Abstract (arXiv)

We study how to improve a frozen pretrained language model with a small amount of adaptive extra computation. A simple approach is to add additional refinement steps on top of the backbone hidden states, but fixed extra refinement can be wasteful: a one-step refinement head may be too weak, while forcing a second full-sequence refinement step everywhere can increase compute without improving transfer. We introduce HALO, a hybrid adaptive latent-refinement method that combines a coarse refinement stage with selective second-stage latent refinement on a subset of tokens chosen by token scoring and monotonic token halting. On the main public benchmark comparison built from MMLU-Pro and GPQA-Diamond, HALO achieves the best overall average among the paper-facing methods, outperforming the frozen backbone, fixed-1, and fixed-2. Internal analysis further shows that HALO reaches nearly the same token-accuracy level as fixed-2 while using fewer average applied refine steps than fixed-1 and far fewer than fixed-2. These results suggest that the key advantage is not simply more refinement, but a better allocation of refinement: HALO achieves the strongest paper-facing result while also using less measured controller compute than either fixed baseline.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 29 of its 30 references that are in the map. First posted on arXiv 3 May 2026.

Builds on

29 of 30 references.

Closest ancestors · 10

Same direction, newest first

  1. 2026Thinking in Latents: Adaptive Anchor Refinement for Implicit Reasoning in LLMs16 Mar
  2. 2026AdaPonderLM: Gated Pondering Language Models with Token-Wise Adaptive Depth2 Mar
  3. 2026LoopFormer: Elastic-Depth Looped Transformers for Latent Reasoning via Shortcut Modulation11 Feb
  4. 2026Pretraining with Token-Level Adaptive Latent Chain-of-Thought9 Feb
  5. 2025Reasoning with Latent Thoughts: On the Power of Looped Transformers24 Feb
  6. 2025Inner Thinking Transformer: Leveraging Dynamic Depth Scaling to Foster Adaptive Internal Thinking19 Feb
  7. 2025Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach7 Feb
  8. 2024Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters6 Aug
  9. 2023Think before you speak: Training Language Models With Pause Tokens3 Oct
  10. 2023Self-Refine: Iterative Refinement with Self-Feedback30 Mar

Other references · 19

  1. 2022Chain-of-Thought Prompting Elicits Reasoning in Large Language Modelscited by 12,459
  2. 2020Measuring Massive Multitask Language Understandingcited by 7,963
  3. 2023GPQA: A Graduate-Level Google-Proof Q&A Benchmarkcited by 2,879
  4. 2024MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmarkcited by 1,848
  5. 2023Enhancing Chat Language Models by Scaling High-quality Instructional Conversationscited by 839
  6. 2024Training Large Language Models to Reason in a Continuous Latent Spacecited by 797
  7. 2024Phi-4 Technical Reportcited by 711
  8. 2018Universal Transformerscited by 647
  9. 2025Phi-4-Mini Technical Report: Compact yet Powerful Multimodal Language Models via Mixture-of-LoRAscited by 521
  10. 2016Adaptive Computation Time for Recurrent Neural Networkscited by 506
  11. 2024LayerSkip: Enabling Early Exit Inference and Self-Speculative Decodingcited by 307
  12. 2024Mixture-of-Depths: Dynamically allocating compute in transformer-based language modelscited by 258
7 more
  1. 2024Lessons from the Trenches on Reproducible Evaluation of Language Modelscited by 179
  2. 2021PonderNet: Learning to Pondercited by 168
  3. 2025Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computationcited by 133
  4. 2025PonderLM: Pretraining Language Models to Ponder in Continuous Spacecited by 44
  5. 2024Early Exit Is a Natural Capability in Transformer-based Models: An Empirical Study on Early Exit without Joint Optimizationcited by 9
  6. 2026Understanding Dynamic Compute Allocation in Recurrent Transformerscited by 6
  7. 2026Tiny Autoregressive Recursive Modelscited by 6

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025Efficient Post-Training Refinement of Latent Reasoning in Large Language Models
  2. 2025MMRefine: Unveiling the Obstacles to Robust Refinement in Multimodal Large Language Models
  3. 2025RefineBench: Evaluating Refinement Capability of Language Models via Checklists
  4. 2025SIGMA: Refining Large Language Model Reasoning via Sibling-Guided Monte Carlo Augmentation
  5. 2025A Sober Look at Progress in Language Model Reasoning: Pitfalls and Paths to Reproducibility
  6. 2025Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning
  7. 2026Enhancing reasoning accuracy in large language models during inference time
  8. 2026Learning to Refine Hidden States for Reliable LLM Reasoning