acceptodds
Under review as a conference paper at ICLR 2027

LatentHalt: Reusing Termination Geometry for Adaptive Latent Reasoning

Abstract

Latent chain-of-thought (CoT) reduces reasoning cost by replacing textual rationales with continuous states, but fixed latent budgets leave the model without a per-problem stopping rule. We introduce , which reuses the termination geometry already present in an explicit-CoT model. In a supervised fine-tuned (SFT) 1B model, last-layer states that predict concentrate in a narrow angular cone, separable from other generation states with held-out AUROC . We freeze the cone's center and quantile radii, then train latent reasoning through a single joint curriculum. An auxiliary decoder supervises the content of each block, while a linear terminal hinge trains the final state of a fully latent sequence to enter the cone. At inference, the same region determines when to stop generating latent blocks and decode the answer, without a separately trained halting head. Across GSM8K-Aug, GSM-Hard, MultiArith, and SVAMP, all problems halt adaptively, using blocks on average. is on par with fixed-length supervised latent CoT on GSM-Hard and improves SVAMP and MultiArith accuracy by and points. It uses fewer estimated reasoning-phase FLOPs than explicit CoT, and its best model comes from a single -epoch run rather than the two-phase, -epoch CoconutSIM-CoT pipeline. The code is available at https://anonymous.4open.science/r/LatentHalt-EA19/.

Primary area:
foundation or frontier models, including LLMs
Keywords:
latent chain-of-thought, continuous reasoning, adaptive halting, hidden-state geometry, termination geometry, adaptive computation, efficient language model 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_eYePOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LatentHalt, an adaptive stopping method for latent chain-of-thought reasoning. It calibrates a termination region from an explicit-CoT model and uses a joint curriculum combining auxiliary step supervision with a geometric stopping loss. During inference, latent generation stops when a block-end state enters this region. Experiments across 1B, 3B, and 8B models report adaptive halting and reduced estimated inference FLOPs on four arithmetic benchmarks.

Critical strengths
  1. Clear motivation and simple design. The paper addresses the practical limitation of fixed latent budgets with a lightweight geometric stopping rule that requires no separately trained halting head.
  2. Useful empirical analysis. The geometry analysis, cross-domain evaluation, and loss ablations help characterize the stability of the stopping signal and its interaction with step supervision.
Critical weaknesses
  1. The baseline comparison is incomplete. Table 1 includes Coconut-based SIM-CoT but omits CODI and CODI+SIM-CoT, while Table 3 includes only Coconut as the competing latent baseline. This leaves the competitiveness of the reported accuracy–compute trade-off unclear.
  2. The advantage of geometric halting over simpler alternatives remains unclear. Appendix C includes fixed one- and ten-block controls, but comparisons with intermediate fixed budgets at comparable compute and a learned halting head under the same joint training are missing.
Questions for authors
  1. Could the authors compare against CODI and CODI+SIM-CoT under matched settings, reporting both accuracy and inference cost?
  2. Does the geometric stopping rule outperform a simple learned halting head and fixed-budget alternatives under the same joint training procedure? Could the no-region-loss ablation also be evaluated with suitable fixed budgets?
  3. Could the authors report measured inference latency and training GPU-hours to establish how the estimated FLOPs and epoch reductions translate into practical efficiency gains?
Reviewer_7qD6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper explores how to adaptively stop reasoning in continuous chain-of-thought tasks. It begins by fitting a fixed angular area to hidden states that anticipate the final token in an explicit CoT model. During a prefix-replacement training phase, an auxiliary decoder keeps an eye on each latent block. Terminal and nonterminal losses then help block-end states decide whether to enter or stay out of this area. Inference is stopped when a state crosses into the region. On four arithmetic problems, the 1B model consistently stops within its allotted time on every instance, averaging 2.43 blocks. It also boosts average accuracy compared to the fixed-length latent models mentioned earlier, while also lowering the estimated number of FLOPs needed to continue. Even though explicit CoT is still much more accurate, experiments with 3B and 8B models also show that halting works reliably.

Critical strengths
  • The stopping rule is simple and inexpensive.
  • The ablations usefully distinguish successful termination from successful reasoning.
Critical weaknesses
  • Adaptive Stopping Evaluation: Unclear how much of LatentHalt’s performance comes from adaptive stopping versus changes in training and rollout length.
  • Stopping Region Benefit: Unclear if inheriting the SFT termination region provides specific benefits or if a generic supervised target would suffice.
  • Comparison with Other Adaptive Stopping Methods (https://arxiv.org/abs/2511.21581): A comparison with a fixed random-center control and a learned linear halting head is needed to understand the impact of inherited geometry versus a generic supervised target.
Questions for authors
  1. Why no learned-halting baseline? A linear probe on block-end states trained to predict the SFT model's own </think> decisions would directly test whether the frozen geometric region adds anything over learning to stop. And given FR-PONDER exists — same problem, learned controller, a year earlier — what does frozen geometry buy you over it, in behavior and in compute?

  2. The 8B Coconut baseline (16.9% GSM8K, below No-CoT) looks collapsed. Can you report a working fixed-budget latent baseline at 8B, or diagnose the collapse? Table 3 as it stands doesn't support the claimed advantage over fixed-length methods.

  3. How much does the termination region depend on the custom <think>/</think> protocol? Does the same concentrated, separable region appear with a standard SFT format (numbered steps, then answer), or a different seed?

Reviewer_NunEOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper finds, in explicit-CoT model, the last layer state before generating </think> token has same direction and lies in a narrow angular cone. Using these feature to train model when to stop thinking, finding the model can be more efficient than coconut while achieve better performance on in-distribution and out-distribution math benchmarks.

Critical strengths
  1. The method is intuitive and well-defined, with clear justification
  2. Model has achieved better performance on math compared with coconut and maintains better efficiency.
Critical weaknesses
  1. Current evaluation focuses on elementary arithmetic problems. Using the model trained on GSM8K-Aug and the same frozen termination region, does LatentHalt maintain a favorable accuracy and compute trade-off on more challenging mathematical reasoning benchmarks?
  2. For explicit-CoT SFT models on non-mathematical tasks such as code generation, do terminal states that predict </think> exhibit similar directional concentration and remain well separated from nonterminal states?
Questions for authors
  1. How sensitive is LatentHalt to the relative weighting of the terminal and nonterminal terms in the region loss? While the paper compares linear and squared terminal penalties, the effect of the weight ratio remains unclear. Could the authors fix the terminal weight and vary the nonterminal weight, including setting it to zero, and report the effects on halting behavior, accuracy, and computational cost?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.