acceptodds
Withdrawn from ICLR 2027

Latent Recirculation: Refining Reasoning through Gated Latent Feedback

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Chain-of-thought (CoT) reasoning improves the performance of language models by making intermediate steps explicit, but decoding these steps autoregressively incurs substantial output overhead. Latent reasoning instead performs intermediate computation in continuous representations, offering a compact alternative to explicit thought generation. However, constructing useful latent updates and supervising their evolution remain challenging. We introduce Latent Recirculation, a lightweight framework that constructs a fixed, question-conditioned latent workspace and refines it through gated deep-to-shallow feedback. The feedback allows the reasoner to revisit the same workspace without increasing its length or generating intermediate text. To guide this refinement, an auxiliary explicit CoT branch provides a reference state during training, and a transition objective aligns latent updates with the displacement toward that reference. Both pretrained backbones remain frozen, with optimization restricted to the latent interface and low-rank adapters on the reasoner. Experiments on symbolic, arithmetic, and commonsense reasoning benchmarks show improved average accuracy across two reasoners, while ablations highlight the importance of combining feedback with transition supervision.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Latent Reasoning, Continuous Thought, Recurrent Reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_iDDSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Latent Recirculation to refine a fixed-length workspace for implicit language-model reasoning. A frozen smaller model produces question-conditioned features, which a trainable projection converts into latent slots for a larger reasoner. One additional pass feeds the reasoner's deep latent states into its first-layer keys and values to revisit the same workspace. A training-only CoT branch supplies a detached reference for directional alignment, while answer supervision trains the latent interface and reasoner adapters. Experiments on symbolic, arithmetic, and commonsense benchmarks show the highest five-task average among the compared methods on both main backbones, with gains varying across tasks.

Critical strengths
  1. The idea is interesting. The fixed workspace and prefill-only feedback provide a concrete way to add latent computation without extending the intermediate sequence or generating a reasoning trace.
  2. The paper is well-written and easy to follow.
Critical weaknesses
  1. Eqs. (6)-(7) constrain the angle of the update but do not establish that the latent state moves closer to a useful reference. An arbitrarily small update or an overshooting update can have the same favorable angle. Gradients also pass through the previous state in the realized transition, while the detached reference changes as the shared adapters are updated. Therefore, reduced alignment loss need not reflect progressive refinement of a fixed reasoning state. The paper does not measure whether improved alignment actually reduces the reference-state discrepancy or improves answer quality along the update.
  2. The no-feedback ablation removes both recirculation and CoT supervision, while every supervised alignment variant retains feedback. This leaves unresolved whether the extra pass is needed to benefit from the reference representation. A single-pass model trained with the same CoT-conditioned reference would isolate that contribution more directly. This distinction matters because feedback without alignment performs worse than no feedback on both backbones, and the full method improves over no feedback by only 2.62 and 1.12 average points.
Questions for authors
  1. How much of the reference displacement comes from the content of the reasoning trace, rather than its added length and the shifted positions of the latent slots? Would a length-matched shuffled-trace control preserve the alignment benefit?
  2. The direct-alignment baseline uses cosine similarity between states, whereas transition alignment compares vectors centered at the previous state. Could the gain come from removing a shared residual component through this centering? How do the two losses compare in gradient scale under the same weight of 0.25?
Reviewer_cnkNOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Latent Recirculation, a method for latent reasoning through refinement of a latent representation of the reasoning trace. Refinement occurs by iteratively processing the latent representation in parallel, conditioned on the prompt and the fixed initial representation (provided by a small student model). During training, an additional loss term optimizes the latent representation to represent explicit text reasoning form. Empirical evaluation on standard math datasets and state tracking tasks show the method is competitive with other latent reasoning methods.

Critical strengths
  1. The paper proposes a concrete and lightweight method for latent reasoning that is more efficient to train than the considered baselines.
Critical weaknesses
  1. My main concern is on the reported numbers, the performance of the suggested method, latent reasoning baselines and CoT model are all very low. As a concrete reference, in Coconut (which is considered here as well) a GPT-2 based 200M model obtains ~43% and 24% accuracy on GSM8k for CoT and Coconut respectively. The authors report much more capable base models that reach considerably lower performance.
Questions for authors
  1. Can you explain the reported performance, specifically the CoT based baseline, e.g. Mistral 7B reports 52% accuracy on standard GSM8K
Reviewer_rXqXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Latent Recirculation, which refines a fixed latent workspace by feeding final-layer states back into first-layer keys and values. A training-only CoT branch supervises the direction of the latent update. The method achieves the highest average accuracy among the evaluated latent baselines, with a small trainable parameter footprint.

Critical strengths
  1. The design is relatively simple: one additional feedback pass, a fixed latent length, and no intermediate text generation at inference.
  2. The ablations show that transition supervision matters. Directional alignment outperforms direct state matching on both main backbones.
  3. The evaluation covers multiple tasks and model families, and includes training cost and inference latency.
Critical weaknesses
  1. The gains over the no-feedback variant are modest: 2.62 points on Llama-2 and 1.12 on Mistral. GSM8K accuracy decreases on both, so the evidence for improved mathematical reasoning is limited.

  2. The no-feedback ablation also removes transition supervision. A no-feedback model with comparable CoT distillation is needed to separate the contribution of feedback from that of additional supervision.

  3. The explicit CoT baseline is zero-shot, while the proposed method receives task-specific training. Although acknowledged, this leaves the practical advantage over equally trained CoT and answer-only baselines unclear.

  4. The novelty is mainly in the combination. Recirculation already introduces normalized deep-to-shallow mixing, while GLR supervises CoT-derived latent updates. LRT(https://openreview.net/pdf?id=CbK7lYbmv8) also uses a lightweight network to generate fixed-length latent reasoning representations in a single pass. The paper should discuss these connections and clarify what its feedback and transition supervision add.

Questions for authors

In addition to previous weakness, what is the latency and memory overhead relative to the no-feedback variant? This comparison would help assess whether the additional pass is worthwhile.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.