acceptodds
Under review as a conference paper at ICLR 2027

LOCKR: A Hidden-State Trajectory-Guided Planner for Detecting and Repairing Stable-but-Wrong Lock-In in Diffusion Language Models

Abstract

Diffusion language models generate text through iterative denoising, exposing intermediate trajectories before final answers are produced. We identify a recurring reasoning failure, stable-but-wrong lock-in, where an answer stabilizes early around an incorrect value while substantial denoising remains. Surface-level decoding signals such as confidence, entropy, margin, and answer stability are insufficient to reliably distinguish correct from erroneous lock-in. We formulate selective reasoning repair as a lightweight test-time planning problem and propose LOCKR, a hidden-state trajectory-guided planner that decides when to allocate additional computation, expands a structured set of targeted repair branches, and selects the most promising continuation using trajectory-aware verification. Across two diffusion language models and three mathematical reasoning benchmarks, hidden-state trajectories consistently outperform surface signals and single hidden snapshots for both wrong-lock-in detection and repair selection. On natural evaluation distributions, LOCKR yields absolute accuracy gains of 2.21–5.37 percentage points across all five evaluated settings, with repair rates ranging from 22% to 41%. These results establish hidden diffusion trajectories as actionable signals for selective test-time reasoning repair.

Primary area:
generative models
Keywords:
diffusion language models, stable-but-wrong lock-in, hidden-state trajectories, test-time planning, reasoning repair, selective inference

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_KF11OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This work argues for the importance of distinguishing the convergence of a discrete diffusion model and its correctness. They coin several concepts such as "stable but wrong" (SBW), and "lock-ins" (to quantify stability, even when the produced answer is wrong). They then propose to evaluate the model's hidden representations with an additional classifier to intervene with repairs, and which measurably increases the performance of the base models at a high computational cost.

Critical strengths
  • This paper presents a relatively simple algorithm/method to measurably improve performance of baseline models.
  • The authors identify an issue that seems important to the diffusion literature (here called “lock-in”).
Critical weaknesses

Weaknesses:

  • There is a lack of clarity in the paper about the nature of the diffusion language model. I originally thought the model was a continuous diffusion language model, but I guess the citations and the models evaluated in Table 9 made it clearer that it was discrete diffusion. I would make this more explicit.
  • The specification of the extraction function, E\mathcal{E}, is not clear where it is introduced: “a deterministic extraction function [yielding] the current normalised answer”. In particular, what is “normalised”, precisely? (I can imagine a few things, but this should be exact immediately.)
  • What is a “Surface-18” vector? I assume that the authors wanted to introduce the nomenclature, but, in which case, it should be written the other way around, unless I am unaware of this specific terminology.
  • The terminology of 3.3 is very hard to read. A lot of vocabulary is implicitly introduced (i.e., without giving the actual meaning of words; e.g., “caliper”, “Surface-18”, “Matching”). Worse, the experimental benchmark should appear directly in 3.3, or else it is not understandable when first reading the paper. Most surprisingly, an explanation is finally provided at the last sentence of the subsection (“This tests whether hidden trajectories…”). I have to admit that I find this section to be a very hard read.
  • What is “(selective) reasoning repair”? Used multiple times, never introduced. What is “prefix hidden trajectories”? In general, “The observed prefix hidden trajectory forms the planning state: a risk model either retains the baseline path or triggers targeted remasking to expand a local counterfactual repair space”, and other sentences are very confusing reads.
  • “We define Ai=(ρi,Ki)\mathcal{A}_i = (\rho_i,K_i), where ρi\rho_i selects an intermediate baseline state and KiK_i is the nominal remasking budget applied to an answer-related region before denoising resumes”: what is ρi\rho_i? a scalar? a function? Likewise: what is KiK_i?
  • Similarly, and there are more examples, but I will end with this last one, H\mathbf{H} is used before it is introduced. I really do not think that giving an overview, that tries to be formal but isn’t, is such a good idea. I think a sort of “linear” build-up/narrative would have made things much clearer.
  • An algorithm would have made the repair procedure much clearer.
  • Overall, while the performance improvements in the experiments are real, I struggle to see the connection to the original motivation of the wrong lock-ins. If I understand the algorithm correctly, ranking/measuring the potentially wrong tokens with an additional head, and branching out into 8 tracks is evidently helpful; the model benefits from some additional parameters certainly too, and from some smarter sampling scheme, which relies on the rich hidden states.
  • The computational cost incurred by the method is large.
Questions for authors
  • Could you produce an algorithmic box for your final method to alleviate any ambiguity about what your method is doing exactly?
  • How does your original understanding of the distinction between convergence and correctness relate to your proposed algorithm? I guess it is obvious in some sense, but I don't see the insight, and find the two parts quite disconnected. Adding a new model which evaluates different branches at inference sounds like a natural way to improve performance, but, again, the connection with the motivating bit is too brittle in my understanding.

I would be happy to revise my score slightly after the clarification of many of the given concepts, a large re-write (which might be out of scope for a rebuttal, but maybe feasible with some edits only), and especially the better establishment of a connection between the motivation and proposed solution -- which I find currently too disconnected.

Reviewer_xcqHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LOCKR studies stable-but-wrong intermediate answers in diffusion language models. It constructs surface-matched correct/wrong lock-in pairs, encodes projected hidden states across four denoising snapshots with a bidirectional GRU, and uses the resulting risk estimate to trigger eight fixed remasking actions. A separate trajectory verifier and gradient-boosted selector choose a repair branch. Experiments cover two model families and three numeric-answer mathematical datasets. Hidden trajectories improve detection and fixed-pool branch selection over surface statistics and validation-selected snapshots; the reported natural-distribution accuracy improvements are 2.21–5.37 percentage points.

Critical strengths
  1. The controlled detection study provides useful evidence that internal states contain correctness information beyond matched observable confidence/stability features. Matching never uses hidden states, its residual imbalance is reported, and the snapshot observation point is selected on validation data. Trajectories outperform the selected snapshot in all five settings, by 1.33–3.61 AUROC points (pp. 4, 7; Appendices B and E.2, pp. 18–21 and 32–33).

  2. The repair evaluation separates candidate availability from selection quality. All fixed-pool selectors see the same branches, LOCKR improves on snapshot selection in every setting, and the oracle/coverage analyses explicitly expose failures of the action bank rather than attributing every error to detection. Pair-level splits and grouping of all derived branches are sensible safeguards (pp. 15–16, Appendix A.3; Tables 24–26, pp. 34–37). The held-out natural-distribution gains also provide evidence of practical benefit within the evaluated protocol (Table 6, p. 9).

Critical weaknesses
  1. The end-to-end comparative evidence does not yet establish the main practical advantage on final outputs. Section 5.1 explicitly defines the controlled benchmark’s 50% baseline using checkpoint correctness rather than correctness of the unmodified completed generation (p. 6), although Section 3.1 allows later answers to change (p. 3). The natural-distribution experiment does evaluate overall accuracy, but Table 6 compares only base decoding and LOCKR, without the competing repair/snapshot methods (p. 9). Thus it does not show whether the controlled comparative advantages persist on actual final answers under the source distribution. Moreover, the source protocol is answer-first prompting with a leading numeric answer, a consequential restriction disclosed in Appendix A.2 (p. 15).

  2. The efficiency evidence is not sufficiently auditable. Base generation is configured for 32 denoising steps (Table 8, p. 14), while every triggered example has eight branches, each configured for 96 continuation steps (Appendix D.1 and Table 16, pp. 26–27); nevertheless, always repairing costs only 5.0x (Table 7, p. 9). Different block schedules or realized continuation lengths could explain this, so this is an unresolved accounting ambiguity rather than a demonstrated numerical error. Appendix A.4 defines cost using unspecified denoising work without giving the executed forward/update counts or a per-action calculation (p. 17). Because the accuracy–compute trade-off is central, this mapping and comparisons at matched realized budgets are necessary to assess the claimed efficiency.

Questions for authors
  1. For the same held-out examples, what are the final-output accuracies of unmodified decoding and each main competing repair method, using the same completed-answer evaluator? Please report these comparisons on the natural distribution at comparable realized budgets, together with its exact sample counts. This would resolve whether the comparative advantage survives beyond the checkpoint-label controlled evaluation.

  2. What exact model executions constitute one unit of denoising work for each backend? Please give the realized forward-pass/active-token counts for base decoding and each of the eight 96-step repair continuations, and show how they yield the 5.0x always-repair cost. Are extra model passes needed to obtain hidden features at inference included?

  3. Does the detection/repair advantage persist with the models’ normal reasoning-before-answer prompts? If this has not been tested, how much of the SBW prevalence and benefit should be interpreted as specific to the answer-first protocol in Appendix A.2?

Reviewer_P6fhOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

dLLMs refine a draft of the whole output, so on a math problem the answer can be read off halfway through denoising. The paper identifies and tackes the recurring reasoning failure of dLLMs, which maens an answer stabilizes early around an incorrect value while substantial denoising remains. Many of these wrong early-settled cases look the similar to the correct ones in terms of confidence, entropy, margin and stability. To ask whether the hidden states know better, the authors pair each wrong early-settled example with a correct one matched on several surface statistics, and train a small recurrent model on compressed hidden states collected across layers and denoising steps to predict whether a settled answer is wrong. LOCKR uses the risk score to trigger repair: re-mask the answer region from a few points of the original run under a fixed menu of settings, re-denoise, score the branches with a second learned model, keep one. The method is evaluated on 2 models and 3 benchmarks.

Critical strengths
  • The surface-matched construction is an interesting idea
Critical weaknesses
  • The phenomenon is likely just a side effect of how the prompt is phrased. The study runs under an answer-first format: outputs begin with #### <answer> and the detector's answer track is the first eight generated tokens. Forcing the answer before any reasoning is the obvious cause of an answer locking in early, and it could possibly depresses base accuracy. Remasking the answer after more of the sequence is denoised is then just letting the model answer after reasoning. The most commonly used prompt for reasoning benchmarks is reason first and answer in the end, but no such baseline is reported, so nothing in the paper separates a real failure mode from a prompt choice, and that is the premise of the whole paper.
  • The practical evidence does not support the claims. All learned components need correctness labels on tens of thousands of generations per model-dataset pair, and gains are reported on held-out slices of training corpora, with no cross-dataset transfer. Commonly used math benchmarks like GSM8K or MATH-500 is suggested to be included. In Table 6 the only comparison is base vs LOCKR, with none of the compute-matched baselines of Table 4 run there. The key claim that hidden trajectories beat surface signals where it matters is untested.
Questions for authors

In addition to addressing the concerns mentioned above, could the authors show how the method works for other widely used dLLMs, such as Dream?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

Dear Reviewers and Area Chair,

Thank you for your detailed feedback and the discussion. We have uploaded a revised manuscript incorporating the clarifications, additional experiments, and evaluation corrections described in our responses. For ease of reference, we summarize the main updates below.

1. Definitions and method protocol (Sections 3–4; Appendices A–D). We clarify answer extraction, normalization, correctness, and Surface-18 matching without hidden-state information. Algorithm 1 explicitly describes the complete-Base protocol, including midpoint-prefix risk estimation, eight fixed repair actions, candidate verification and selection, and acceptance or fallback to completed Base. The revised text also distinguishes the detector, verifier, and selector.

2. Final-output evaluation (Sections 5.4–5.6; Appendix F). We distinguish the matched benchmark’s checkpoint labels from completed-Base accuracy and re-evaluate all policies on their actual returned answers across 38,416 questions. We add seven-method comparisons on five new natural-distribution holdouts, each with 10,000 test questions and a separate 2,000-question validation set. The full results include ineligible questions, Rescue/Damage counts, routing rates, computational costs, and paired confidence intervals for the natural-distribution comparisons.

3. Computation and timing (Appendices A.4, D, and G.3). We explain the historical analytic 5× convention and distinguish accumulated forward-token work, profile-weighted estimates, and elapsed time. The new accounting includes individual repair actions, hidden-state extraction, required replay, candidate generation, auxiliary scoring, and routing. We also separate the nominal action bank from configuration-specific profiling and clarify that savings over always-repair do not imply a speedup over Base.

4. Prompt controls, Dream, and frozen transfer (Sections 5.7–5.9; Appendices I–K). We add answer-first/reasoning-first comparisons, late answer regeneration, and RF-LOCKR-pos, together with answer-visibility and eligibility statistics. Frozen GSM8K evaluation covers all 1,319 test questions for DG, LLaDA, and Dream without target training or calibration. Dream uses its own source-trained components, and its RF Base check is explicitly distinguished from a full RF planner evaluation.

5. Component and supervision controls (Sections 5.10–5.11; Appendices G.1 and L). We add independent trigger comparisons, clean selection comparisons with fixed repair IDs and candidates, matched-capacity and training controls, retrained shuffled-order controls, and error-enrichment diagnostics. Complete-pipeline supervision curves cover four training fractions and three seeds, with explicit supervision counts and preparation and retraining cost scopes.

6. Conclusions and limitations. We update the abstract, introduction, conclusion, limitations, captions, and cross-references to reflect the revised evidence. We retain uncertain and unfavorable comparisons and qualify claims concerning prompt dependence, shared data lineage, candidate coverage, temporal order, supervision, and computational overhead.

We appreciate the reviewers’ and area chair’s time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.