acceptodds
Under review as a conference paper at ICLR 2027

LatentWeave: Learning High-Value Thought for Efficient Visual Reasoning

Abstract

Scaling reasoning increasingly relies on longer chains of thought. Because natural language expresses both reasoning progress and elaboration of existing thought, scaling explicit reasoning does not necessarily constitute more intelligence. We hypothesize that reasoning capability depends on high-value thought and its coherent organization. Motivated by this hypothesis, we introduce \method, combining rubric-guided reinforcement learning to retain high-value explicit reasoning with interleaved latent updates to maintain reasoning continuity. Experiments on six benchmarks covering hard multimodal, vision-dependent, and long-chain reasoning establish an empirical efficiency–accuracy Pareto frontier. Relative to the Qwen3-VL-8B-Thinking baseline, \method achieves reasoning efficiency with 1.86% relative accuracy loss, and efficiency and 3.84% loss. This Pareto frontier supports flexible model selection on varied reasoning budget and performance requirements.

Primary area:
reinforcement learning
Keywords:
VLM, efficient reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_32G1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how to shorten explicit reasoning in vision-language models while retaining problem-solving ability. LatentWeave jointly trains two components. First, a rubric evaluates complete reasoning trajectories for necessary stages, visible reasoning structure, compression safety, answer grounding, and related criteria; this feedback is combined with correctness and length signals for reinforcement learning. Second, the model selectively adds continuous residual updates to the input embeddings of subsequent ordinary tokens. These updates can access an additional visual feature channel, are activated by a routing module, and introduce no extra output tokens or KV sequence positions.

The evaluation covers six visual reasoning benchmarks. Relative to Qwen3-VL-8B-Thinking, two principal checkpoints reduce mean visible output length from 3,649 tokens to 1,252 and 660 tokens, corresponding to 2.92× and 5.53× compression. Macro-average accuracy decreases from 74.69% to 73.30% and 71.82%, respectively, or by 1.39 and 2.87 percentage points. The paper also reports same-backbone training trajectories, a text-only rubric control, repeated-sampling evaluation on MathV-G100, inference-time component interventions, and reasoning-trace case studies. The compression ratios concern visible output length, rather than measured end-to-end speedups.

Critical strengths

1. The paper provides an empirically supported approach to visual reasoning compression, with evidence beyond individual examples. Table 1 demonstrates substantial output-length reductions across six benchmarks and explicitly reports their accuracy costs: 1.39 percentage points in macro-average accuracy at 2.92× compression, and 2.87 points at 5.53× compression. The paper also exposes differences across tasks rather than reporting only an aggregate. These evaluated operating points are useful for applications with constrained visual reasoning budgets and are the main reason for my positive recommendation.

2. The joint design is technically meaningful and is evaluated from complementary perspectives. Applying continuous residuals at existing token positions allows explicit reasoning selection, latent state updates, and visual information access to be learned within the same generation process. Appendix Figure 5 includes a text-only comparison across all six benchmarks; Table 2 examines both single-sample accuracy and repeated-sampling solution coverage; and Table 3 tests the trained model’s dependence on the visual channel, latent updates, and routing. These experiments do not completely isolate the causal contribution of each component, but together they provide useful evidence for the joint system beyond compression ratios or selected successful traces.

Critical weaknesses

1. The methodological novelty appears incremental, with the contribution primarily in a targeted integration of established ideas. This is my main reservation. RL-based length control, rubric-based rewards, integration of continuous representations into textual generation, and routing between discrete and latent computation have precedents in L1, RaR, HRPO, and ThinkRouter, respectively. LatentWeave has concrete technical distinctions: bounded additive residuals derived from hidden states, visual evidence access, and joint training of semantic routing and frequency regulation with a compression-oriented rubric. These distinctions deserve credit. Nevertheless, the contribution currently appears strongest as a system integration and empirical advance for visual reasoning compression, with a more limited standalone methodological advance. The paper should more clearly establish which distinctions address limitations of existing approaches and why they matter for the reported gains. 2. The incremental benefit of the added design over simpler alternatives is not yet sufficiently established. When the contribution rests primarily on a combination of methods, the benefit of that combination is especially important. Appendix Figure 5 shows that Pseudo action and LatentWeave both reach 5.53× compression, with relative macro-accuracy losses of 4.67% and 3.84%, corresponding to an advantage of approximately 0.62 absolute percentage points for LatentWeave. This is positive evidence, but the large compression ratio is already attainable with this text-only approach. Moreover, the curves represent different training lineages, limiting assessment of the stable incremental benefit. The two comparisons in Table 2 do not simultaneously match training budget and output length, and a Standard RL plateau under one recipe does not establish a well-optimized text-only accuracy–length frontier. The results therefore support the effectiveness of the complete system more strongly than they establish the value of its most distinctive design choices. This is a concern about relative contribution, not a claim that the 0.62-point gain is worthless or statistically insignificant.

Questions for authors

Additional experimental comparison with the closest prior methods. Section 2 discusses HRPO and ThinkRouter, but neither has results in Table 1 or the appendices. Could the authors prioritize adapting at least one of these methods to the same Qwen3-VL-8B-Thinking backbone and compare its accuracy–inference-cost tradeoff with LatentWeave on the same visual benchmarks and evaluation protocol? Please report accuracy, visible output tokens, and additional latent computation at comparable budgets; for methods requiring training, use the same training data and comparable training budgets. These results would directly test the practical gain over prior hybrid latent reasoning methods. This is a request for an additional baseline experiment.

1 comment

We thank Reviewer 32G1 for recognizing the empirical value of LatentWeave. We clarify the contribution and provide revised controls and a same-backbone ThinkRouter comparison.

Relation to prior work and novelty

LatentWeave learns to interleave concise explicit reasoning with visually informed latent updates. Each Action both updates the model state and marks a boundary in the visible reasoning. Rubric feedback assesses the reasoning and its continuity across these boundaries, jointly training text generation and the boundary policy (Submission §2--3; Revised §2--3).

Correctness and length rewards evaluate the answer and its cost, but do not explicitly assess whether the reasoning provides sufficient premises, valid deductions, and necessary checks. A rubric supplies that assessment, yet alone remains text-only compression. In LatentWeave, the learned boundary links the reasoning being assessed to the state update being executed: explicit reasoning and latent updates develop together during training, rather than replacing selected spans of a pre-existing CoT. The feedback remains trajectory-level, not a local causal score for each Action.

ActionNet makes this boundary computational rather than merely annotational: it injects a bounded, visually informed residual at an existing token position, without extra tokens or KV positions. The key empirical question is therefore whether the benefit comes only from showing boundaries to the judge, or also from learning where to place them and executing latent updates there. Revised Table 2 and Figure 5 address this through the controls below.

A complementary contribution is compression-aware frequency regulation. As reasoning traces shorten, fewer positions remain available for latent actions. We adapt the distance-prior scale when mean Action counts fall below the target, preserving execution opportunities while leaving context-dependent placement to the learned semantic router. The fixed-versus-adaptive comparison in Revised Table 2 and Figure 5 isolates this design: adaptive scaling produces shorter outputs at both reported checkpoints, with an accuracy trade-off at step 300 and similar accuracy at step 750.

Incremental benefit over simpler alternatives

Revised Table 2 and Figure 5 test random boundaries, learned boundaries, and latent execution under the same revised training and evaluation settings. The following comparison uses step 750 throughout; Random gate corresponds to the submission's Pseudo action.

SettingBoundary / updateAccuracy (%)Tokens
Random gateRandom / none72.03602
Learned gateLearned / none72.48683
LatentWeaveLearned / visual latent update72.78612

Learned boundaries improve accuracy over random markers but use more tokens. Executing visual latent updates improves accuracy further while shortening the trace: +0.30 pp and 683 → 612 tokens relative to Learned gate. Relative to Random gate, the full model gains 0.75 pp at nearly the same length. Thus, judge-visible segmentation alone does not account for the full model's gain.

Two additional controls support this interpretation (Revised Table 2): at step 300, LatentWeave exceeds boundary-free Rubric RL by 0.53 pp at nearly matched length (1,137 vs. 1,148 tokens); at step 750, adding Visual KV improves accuracy by 0.16 pp and reduces length from 731 to 612 tokens. Figure 5 and Appendix E.1 report the full trajectories.

Requested baseline and computation comparison

We add ThinkRouter on the same Qwen3-VL-8B-Thinking backbone and six benchmarks (Revised Table 1), using temperature 0.6, default threshold 0.7, and top-10 embedding mixing without per-benchmark tuning.

MetricBaselineThinkRouterLatentWeave-300LatentWeave-750
Six-benchmark accuracy (%)75.6874.8974.5472.78
Mean output tokens3,8273,5501,137612
Token compression1.00×1.08×3.37×6.26×
MathVision amortized overhead0≤0.00050%0.47%0.54%

For computation accounting, ThinkRouter uses soft routing on 90.61% of thinking steps (step-weighted; MathVision: 90.90%). This measured frequency weights the embedding-mixing cost in its amortized overhead estimate.

Overhead follows Revised Appendix C.3's per-output-token accounting. ThinkRouter's bound is 0.9090×0.00008192/15.136×100%0.9090\times0.00008192/15.136\times100\%; answer tokens only lower it. LatentWeave assumes 1,024 visual tokens. These are arithmetic estimates, not latency measurements. ThinkRouter is training-free; its default operating point is not token-budget matched.

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

This paper proposes a latent reasoning framework for efficient visual reasoning named LatentWeave. Specifically, LatentWeave consists of three core components: the Rubric RL module, the Interleaved Latent Actions module, and the Semantic Switching with Frequency Regulation module. The Rubric RL module encourages concise and grounded explicit reasoning through a six-criterion rubric and a correctness-tiered reward scheme. The Interleaved Latent Actions module supports reasoning continuity via bounded residual updates. Finally, the Semantic Switching with Frequency Regulation module adaptively selects latent update positions using a Bernoulli policy and multiscale distance and token-entropy priors.

The experimental evaluation in this paper assesses the effectiveness of LatentWeave across multiple benchmark datasets, including MMMU, MathVision, MMStar, RealWorldQA, M3CoT, and ScienceQA. Results demonstrate that LatentWeave achieves substantially reduced output lengths while maintaining competitive average accuracy.

Critical strengths
  • The idea of using rubric-guided RL to learn which reasoning steps to retain and Latent Actions to support continued reasoning after compression is considered novel.

  • The experiments support the preservation of reasoning capabilities with shorter outputs, although aggressive compression still degrades problem-solving performance

Critical weaknesses
  • The authors state that their goal is to learn which explicit reasoning steps are worth retaining and how latent updates can preserve the information needed for subsequent reasoning when some text is omitted. However, the rubric judge only observes the image, question, visible reasoning, answer, and action positions. It can only assess whether an Action occurs at a location that appears suitable for an update. Even if this allows the judge to assess whether a reasoning segment is grounded, can it reliably determine whether that segment is useful for subsequent reasoning?

  • The correspondence between the rubric used to assess visible reasoning and information retention in the latent states is unclear. The optimization objective does not directly constrain the relationship between the omitted information and the latent updates. There is no evidence that the omitted CoT is fully encoded in the latent representations.

  • I notice that the same trajectory advantage is assigned to all text-token and mode decisions. Does this create a mismatch between the granularity of the learning signal and the decisions? A successful trajectory does not imply that every Action within it is useful; an unnecessary or erroneous action may be reinforced simply because the trajectory ultimately produces a correct answer. Conversely, even if a trajectory contains only an isolated arithmetic error at the final step, the entire trajectory may receive a poor evaluation.

  • Detached replay truncates some gradients across actions, which is at odds with the goal of maintaining long-term state continuity. During actual generation, Actions have recursive dependencies, but fixing the queries during replay omits part of these dependencies from the gradient computation. Although direct gradients with respect to the current ActionNet parameters are preserved, the path through which an earlier residual changes a later query and in turn changes a later residual is removed. Suppose the role of an early Action is not to immediately affect the next one or two words, but to organize an intermediate relation into a representation that subsequent actions can more readily access. Would the proposed replay strategy fail to capture part of this dependency?

Questions for authors
  • An intuitive question that captures the routing objective at a given state is: given the same prefix, does executing this Action improve subsequent outcomes compared with not executing it? Why does the method not estimate this difference directly, instead relying on the router to learn indirectly from trajectory-level statistical associations?
  • As shown in the experiments, omitting intermediate reasoning text incurs greater costs for problems requiring precise visual premises and complex, interdependent deductions. Does this suggest that longer reasoning may still be more valuable than “high-value thought”?
1 comment

We thank Reviewer fjCh for recognizing the novelty of rubric-guided learning with latent continuation. We address the supervision and gradient approximations, and provide a same-prefix intervention.

Rubric supervision and credit assignment: a cost--granularity trade-off

The rubric assesses visible reasoning dependencies, not each Action's causal contribution. Beyond correctness and length, it evaluates grounded premises, valid deductions, necessary checks, auditability, and semantic compression, credits consequential corrections, and retains reliability caps (Revised §3.3 and Appendix A). It can assess whether later visible deductions follow from earlier premises, but cannot verify hidden-state contents or an Action's necessity.

Explicit reasoning and latent updates are learned jointly, rather than by replacing parts of a completed verbal solution. The objective is effective continuation, not complete encoding of omitted CoT spans. Figure 6 illustrates retained premises, deductions, and corrections, with case provenance in Appendix E.4; Table 2 and Figure 5 test policy-level utility.

Shared GRPO advantages do permit coarse attribution: unnecessary Actions can benefit from successful trajectories, while useful decisions can be penalized after a late error. This is a deliberate cost--granularity trade-off. Reliable process rewards for free-form multimodal reasoning require fine-grained supervision and validation across variable steps and solution paths. Moreover, online reward latency lies on the critical path of each on-policy update; finer-grained judging can reduce training throughput, not just increase API cost. We therefore use one rubric assessment per trajectory and separately test the first Action's contribution below.

Detached replay and long-range dependencies

Detachment removes the recursive Action-to-Action gradient, not all learning from later tokens (Revised §3.2 and Appendix C.2).

DependencyTreatment
Early residual → later token likelihoods through the decoderGradient retained
Early residual → later recorded query → later residualGradient truncated
Earlier Actions affecting later states during generationForward dependency retained

An early Action thus learns from later token losses beyond its immediate neighbors, although preparing information specifically for a later Action involves the truncated path. Exact recursive differentiation requires retaining or recomputing the graph, trading activation memory and feasible batch size against sequential work; fixed-query replay instead enables parallel ActionNet evaluation. A related trade-off appears in SWITCH (Appendix A; Algorithm 2), which runs latent segments under no_grad and backpropagates through text segments to reduce peak activation memory. The two methods use different approximations to balance recursive gradient computation against training scalability.

Same-prefix counterfactuals and compression limits

We now test execute-versus-skip from the same prefix. On MathVision-100 with LatentWeave step 750, we branch at the first Action decision and force that Action either to be skipped or executed, then resume normal routing in both branches. Each condition uses eight completions per question under the same Pass@1/4/8 protocol.

First Action interventionPass@1Pass@4Pass@8Mean output tokens
Skip55.75%70.20%76.00%1,798
Execute57.13%72.37%78.00%1,835

Executing the first Action improves Pass@1/4/8 by 1.38/2.17/2.00 percentage points, with 2.1% longer outputs. This supports the first Action's utility on this set under subsequent normal routing. The intervention is an evaluation diagnostic; training uses trajectory-level outcomes rather than a counterfactual estimator because paired continuations add generation cost.

High-value reasoning need not be short: necessary visual premises and interdependent deductions should remain explicit, even in long traces. The larger MMMU and MathVision accuracy losses at step 750 (Revised Table 1) motivate task- and budget-dependent compression, not unconditional length minimization.

Reviewer_RtD6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper compresses the visible chain of thought of Qwen3-VL-8B-Thinking using two components:

  • Rubric-based RL objective: an LLM judge scores six fields, the reward is stratified by correctness, and a log-length correction is applied to correct samples.
  • Interleaved Latent Actions: a ModeSwitcher decides when an ActionNet adds a bounded, RMS-scaled residual to the next token's embedding, attending to layer-0 visual keys/values, with no extra tokens or KV positions.

After joint GRPO training on ~18k VLAA-RL examples, the step-300 checkpoint uses 1,252 tokens vs. 3,649 for the base model across six benchmarks, at a 1.86% relative accuracy loss.

Critical strengths

A practical problem with a careful design. The latent residual is bounded and needs no extra tokens or KV positions. The zero-initialized output projection means training starts from the base model's behaviour, and the gradient boundaries during replay are clearly explained.

The compression is substantial (1,252 vs. 3,649 tokens at a 1.86% relative loss), and the per-task breakdown usefully shows that the loss is concentrated in hard multimodal reasoning.

Transparent reporting. The paper is upfront that it measures visible tokens, not wall-clock speed (§3.4, App. H). It also gives a deterministic rule for building MathV-G100 (D.1), full hyperparameters, and a FLOPs derivation that I re-computed and found essentially correct.

Critical weaknesses

1. The efficiency claims would benefit from measured cost. The evidence is the visible-token ratio plus a FLOPs estimate that excludes self-attention, prefill and vision encoding (C.3). As §3.4 notes, this is "visible-token compression, not a measured wall-clock speedup."

At the same time, the method adds per-step work: the visual keys/values are re-projected at every Action, and top-20 probabilities and an entropy are computed at every eligible position. Without latency measurements, the net efficiency gain is hard to judge.

The "Pareto frontier" claim is also not yet supported by Fig. 1:

  • It rests on two points, and in Fig. 1's own coordinates OneThinker (3.55×, 1.36% loss) dominates step 300 (2.91×, 1.86%).
  • The ratios use different references: Thinking (3,649), Instruct (4,995), and Qwen2.5-VL (312).
  • The step-750 checkpoint behind the 5.53× headline is not in Table 5, which stops at step 400.

2. The contributions of the two components are hard to separate. The only text-only control, Pseudo action, is run only on the 100-question MathV-G100. It also disables the "latent-specific structure and count rewards" (A.2), so it differs from LatentWeave in its reward as well as its latent pathway.

At step 300, Pseudo action is actually shorter than LatentWeave (2,927 vs. 3,357 tokens), which suggests that much of the compression may come from the rubric reward. The key Table 2 differences are small (0.38 points of pass@1, six questions of pass@8) and come from single runs without intervals.

To disentangle them, the following on all six benchmarks would help:

  • rubric RL without the latent pathway;
  • the latent pathway with the Standard RL reward;
  • a controlled ablation of frequency regulation (B.3 notes that the Fig. 3 run can't serve as one).
Questions for authors
  1. Why does Table 5 stop at step 400 when Table 1 reports step 750? Is step 750 from the same run, and what are its MathV-G100 pass@k, token count and Action count?
  2. Could you report rubric RL without the latent pathway, and the latent pathway with the Standard RL reward, on all six benchmarks at matched token budgets? What exactly are the "latent-specific structure rewards" that are disabled for Pseudo action?
  3. Could you report end-to-end latency and throughput on the same inference stack and batch size, including the cost of the repeated visual projection and the entropy computation? If that is not possible, would you consider "token compression" instead of "efficiency" in the title and abstract?
  4. Could all methods in Fig. 1 be normalized to a single reference? Multi-seed or bootstrap intervals on MathV-G100, plus Thinking and Standard RL pass@k on that set, would also help.
1 comment

We thank Reviewer RtD6 for the careful assessment.

Efficiency claims and reference normalization

LatentWeave achieves 6.26× token compression and 5.45× measured decoding speedup (Revised Tables 1 and 3, §4.6; supplementing Submission §3.4).

MetricBaselineLatentWeave-750Relative to baseline
Six-benchmark mean output tokens3,8276126.26× compression
MathVision-100 throughput (token/s)152.12149.1698.1% retained
Full-suite decoding time (min)371.9568.265.45× speedup

Throughput uses the same vLLM stack, batch size 1. Full-suite time is accumulated separately during evaluation. Both exclude prefill, loading, initialization, and judging—not end-to-end latency. Decoding includes routing, entropy, and latent updates. Top-k probabilities are shared between routing and entropy; caching native Visual KV does not eliminate the per-Action adapter projections counted in Revised Appendix C.3.

OneThinker does not dominate the revised accuracy--length results. Revised Figure 1 and Appendix E.1--E.2 replace the submission comparison with corrected metrics, six checkpoints, and training-exposure context:

ModelAccuracy (%)TokensPost-training exposure
Qwen3-VL-8B-Thinking75.683,827Reference
Qwen3-VL-8B-Instruct71.461,523Reference
OneThinker71.36547≈5.14M
LatentWeave-30074.541,13776.8k
LatentWeave-75072.78612192k

OneThinker is shorter but less accurate than LatentWeave. Its reference is Qwen3-VL-8B-Instruct, whereas LatentWeave uses Qwen3-VL-8B-Thinking. Figure 1(a) labels these backbone-specific normalizations; the absolute values above avoid mixing those ratios. Figure 1(b) counts SFT presentations and RL completions, not training FLOPs.

Step-750 results and the former Table 5 issue

Submission Tables 1 and 5 used the same run; evaluation budget limited the diagnostic analysis to step 400. The revised run supersedes these results. Revised Appendix E.1 covers six full-suite checkpoints through step 932; Appendix E.3 covers 19 MathVision-100 checkpoints, including step 750, with the requested Pass@1/4/8, token counts, and Action statistics.

Component disentanglement

Submission Pseudo action (Appendix A.2) corresponds to revised Random gate: text-only rubric RL with random markers shown only to the judge, at 20--100-token intervals. This is the only difference; neither uses latent-specific structure/count rewards. Empirically, shorter marker intervals led to faster output-length reduction during training, but lower accuracy at matched output lengths. This preliminary observation concerns judge-input markers, not model execution frequency; the full marker-interval ablation is not yet complete.

Revised Table 2 and Figure 5 cover all six benchmarks under shared training and evaluation settings:

Requested controlRevised setting
Rubric RL without latent updatesLearned gate: residuals fixed to zero
Latent updates with Standard RL rewardLatentWeave w/o rubric: correctness + length
Frequency regulationw/o Visual KV: fixed vs. adaptive prior scale

The frequency ablation holds Visual KV off and all other components fixed (Revised Table 2, Figure 5). Fixed scale retains its initial value; adaptive scale increases when mean Action counts fall below the target.

StepFixed: accuracy / tokensAdaptive: accuracy / tokens
30074.78% / 1,53474.33% / 1,407
75072.56% / 74472.62% / 731

Adaptive scaling produces shorter outputs at both checkpoints, with an accuracy cost at step 300 and similar accuracy at step 750. This tests adaptive scaling against a fixed prior, not removal of the prior. Figure 3(b)'s historical staged-scale probe (Appendix B.3) motivates the mechanism but is not this controlled ablation.

Selected gains over text-only controls (Revised Table 2):

Control → LatentWeaveAccuracy changeOutput tokens
Rubric RL, step 300+0.53 pp1,148 → 1,137
Learned gate, step 750+0.30 pp683 → 612

The first comparison nearly matches token budgets. Other common-step comparisons have different lengths; Revised Figure 5 and Appendix E.1 report their full trajectories.

Pass@k and multi-seed evaluation

We report Pass@1/4/8 from eight completions per question to assess solution coverage across sampling budgets. Submission Figure 3(a,b) reports Standard RL's Pass@k and output length on MathV-G100; revised Figure 3(a) provides the updated trajectory. Multi-seed evaluation of Thinking and Standard RL under the revised setting is not yet complete. This is separate from the eight-completion Pass@k evaluation. Five of the six Thinking-reference accuracy scores are official; ScienceQA is evaluated locally.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Note for Paper Revision

We thank the reviewers for their careful reading and constructive feedback. The revised paper updates the evaluation protocol and metric aggregation, refines the curated training distribution and rubric, and reruns full-suite component controls under the revised setting. LatentWeave reaches 3.37× and 6.26× token efficiency at checkpoints 300 and 750, with 1.51% and 3.83% relative accuracy loss. For the full six-benchmark suite, pure decoding time is 371.95 minutes for the official baseline, 215.82 minutes for Standard RL, and 68.26 minutes for LatentWeave at checkpoint 750. The backbone and six evaluation benchmarks are unchanged; evaluation sampling changes are specified below.

ScienceQA evaluation and metric aggregation. We found an indexing mismatch between our previous cached ScienceQA questions and answers. The mismatch affected 249 cases, so the ScienceQA evaluation was recomputed after restoring the question--answer alignment. We also corrected the aggregation convention: token counts are now averaged within each benchmark and then combined with equal weight across the six benchmarks, matching the accuracy aggregation, rather than using a pooled per-case mean. The corrected ScienceQA results and aligned token means are used consistently in the revised Table 1, Table 2, the full checkpoint tables in Appendix E, and all reported six-benchmark averages.

Training-distribution refinement. Our analysis showed that part of the original training mixture was relatively easy for the model and was disproportionately mathematical. The mathematical portion was reduced from 49.5% to 42.8% by filtering relatively simple math examples. This produces a more balanced curated VLAA-RL training set of approximately 15k examples. The goal was to make training more efficient and make the resulting accuracy--compression analysis less dominated by easy mathematical cases.

Rubric Refinement. We identified three connected issues during revision. First, the submission's six-dimensional reward has a bottleneck structure: in submission §3.1 and Eq. (1), VB, CS, AG, and the integrity cap enter through a minimum, so a valid self-correction after a latent transition can receive incomplete credit. Second, the submission rubric and wording can give the impression that a latent Action is expected to carry content that would otherwise appear as an omitted CoT span, even though our primary objective is to assess the quality of the retained reasoning trace. Third, these two issues make it harder to distinguish functional latent continuation from verbal-span replacement. We thank the reviewer for raising this interpretation question. The revised rubric in §3.3 reorganizes the assessment around reasoning quality and retains explicit checks on necessary intermediate conclusions and transitions. Revised Figure 6 provides concrete cases, with provenance and excerpt conventions in Appendix E.4; Table 2 and Figure 5 report matched controls.

Submission version, six-dimensional rubric

DimensionIntended rolePotential issue exposed by the case analysis
Semantic-stage coverage (SC)Cover necessary setup, operations, checkpoints, and commitmentDoes not directly credit a concise correction
Visible-backbone quality (VB)Assess the explicit plan, evidence, and transitionsA latent transition can make a valid correction look discontinuous
Latent-operation alignment (LA)Assess whether Action positions support plausible updatesCan shift attention toward boundary placement rather than reasoning quality
Checkpoint grounding (CG)Require concrete results after ActionsA hidden intermediate result may appear unsupported
Compression safety (CS)Check that necessary premises survive without unsupported jumpsCan penalize concise recovery after a latent transition
Answer grounding (AG)Check whether visible reasoning supports the answerA prior visible error can lower support despite a later correction

Revised version, five-dimensional rubric

DimensionEvaluation focus
Premise Grounding (PG)Support for the facts used
Derivation Validity (DV)Valid inference and resolved contradictions
Verification Sufficiency (VS)Coverage of material risks
Trajectory Auditability (TA)Recoverable evidence and intermediate conclusions
Semantic Compression Efficiency (SE)Whether further shortening would sacrifice grounding, inference, risk coverage, or correction
AuthorsOpenReview ↗

Detailed Revision Checklist

Evaluation and data

  • ScienceQA correction. The submission used a cached question--answer mapping with 249 index mismatches. We realigned the cache and recomputed every locally evaluated ScienceQA entry, including the locally measured related-work references and our models. The corrected values propagate to revised Table 1, Table 2, and the full-suite checkpoint and control tables in Appendix E.1.
  • Training distribution. The revised paper uses the curated VLAA-RL training set of approximately 15k examples. We reduced the mathematical share from 49.5% to 42.8% by filtering relatively simple math examples, making the training mixture less dominated by easy mathematical cases.
  • Metric aggregation. The submission did not use the same aggregation convention for accuracy and token length. The revised paper first averages both quantities within each benchmark and then gives the six benchmarks equal weight, so the token statistic follows the same macro-benchmark weighting as accuracy.
  • Evaluation sampling alignment. We aligned evaluation sampling with the settings used for the baseline's official results: temperature 1.0 and presence penalty 0.0, replacing the previous 0.7 and 1.5. Top-p 0.95, top-k 20, and the 40,960-token output cap are unchanged (Revised Appendix D). ThinkRouter uses its separately specified temperature of 0.6.

Training objective and controls

  • Rubric definition. The submission used a six-dimensional rubric. The revised paper uses five criteria in §3.3 and Appendix A, with the scoring equation and reward construction updated to match the revised interpretation of retained reasoning and latent continuation.
  • Full-suite controls. Revised Table 2 and Figure 5 evaluate the major components under the revised training-data and rubric setting. The controls separate rubric feedback, random versus learned boundaries, latent updates, Visual KV, and fixed versus adaptive prior scaling.
  • Frequency mechanism. The frequency-prior equations and semantic routing mechanism are unchanged from the submission. The revision distinguishes fixed from adaptive prior scaling in the controlled comparison; Appendix B.2--B.3 gives the analytic definitions and provenance of the separate historical staged-scale probe.

Checkpoint and case analyses

  • Full checkpoint coverage. Revised Appendix E.1 reports the full-suite checkpoint and training-control results through step 932. Appendix E.3 reports the MathVision-100 trajectory. Figure 3(a) updates the Standard RL length-plateau analysis; Figure 3(b) retains the historical staged-scale probe documented in Appendix B.3.
  • Reasoning cases. Revised Figure 6 shows retained premises, deductions, corrections, and boundary conditions. Appendix E.4 identifies the paired responses and excerpt conventions. These cases make the revised rubric's treatment of self-correction and latent transitions explicit.

Efficiency evidence

  • Runtime measurements. Revised Table 3 adds controlled MathVision-100 decoding throughput using the same vLLM stack with batch size 1, and separately accumulated full-suite decoding time. Both exclude prefill, model loading, initialization, and answer judging.
  • Main comparison. Revised Table 1 reports the corrected six-benchmark results for Qwen3-VL-8B-Thinking, Standard RL, and LatentWeave at steps 300 and 750. It is the main accuracy--length comparison, while Table 3 supplies the separate system-level measurements.
AuthorsOpenReview ↗

Detailed Revision Checklist (cont.)

Major figure and table changes

ChangeRevised-version updateEvidence or issue addressed
Figure 1(a)Reorganized the accuracy--efficiency comparison around each method's corresponding backbone reference and added the common Qwen3-VL-8B-Thinking interpretation in the surrounding text.Makes cross-backbone normalization explicit and isolates the same-backbone comparison requested by RtD6 and 32G1.
Figure 1(b)Added a post-training exposure and absolute accuracy frontier comparison for the selected LatentWeave checkpoints, using the revised trajectory counts and the corresponding external-method evaluations.Adds training-budget context to the efficiency frontier and makes the comparison broader than output length alone.
Figure 2Added a method overview that presents the latent interface, fixed-query parallel replay, and the joint correctness--rubric--length objective. The submission's frequency-prior figure is now Figure 4.Gives the method, training pathway, and reward signal a single structural overview before the detailed mechanism figures.
Figure 3Updated the length-saturation and manual-recovery presentation. Panel (a) reports the revised training behavior between step 600--932, and panel (b) reuses old manual control experiments.Clarifies the motivation for adaptive frequency regulation and the provenance of the manual intervention.
Figure 5Replaced the limited submission's component evidence with revised full-suite training controls covering rubric feedback, random and learned gates, latent updates, Visual KV, and fixed versus adaptive prior scaling.Addresses the request for matched component comparisons and clarifies what each control changes.
Figure 6Reorganized the qualitative compression cases around retained premises, deductions, corrections, and boundary conditions at steps 300 and 750.Provides concrete evidence for functional reasoning retention and self-correction after compression.
Table 1Recomputed the main six-benchmark results after the ScienceQA correction and revised training setting, with common Qwen3-VL-8B-Thinking reference points for the main comparison.Updates the primary accuracy--length evidence and makes the reference convention explicit.
Table 2Added full-suite controls at steps 300 and 750 for Standard RL, Rubric RL, Random gate, Learned gate, LatentWeave, and Visual-KV variants.Separates rubric, boundary, latent, visual-access, and frequency contributions.
Table 3Added measured MathVision-100 decoding throughput and pure full-suite decoding time, excluding prefill, model loading, initialization, and answer judging.Complements visible-token compression with reference system-level measurements.
Appendix E.1Added complete full-suite checkpoint and training-control results through step 932.Resolves the separation between the submission's step-400 diagnostic and step-750 main result.
Appendix E.3Added the complete MathVision-100 trajectory with Pass@1/4/8, output length, and latent-action statistics.Extends the training-dynamics analysis beyond the submission's partial trajectory.
Appendix E.4Records case provenance and excerpt conventions for Figure 6.Identifies the questions and checkpoints used in the paired examples.

Presentation and organization changes

ChangeRevised-version updatePurpose
Main-text narrativeMoved the length-saturation phenomenon before the frequency-control, then placed the full training-control analysis in the experiments section.Establishes the observed compression plateau before introducing adaptive frequency regulation and keeps empirical controls after the method.
Figure sequenceUses Figure 2 for the method overview, Figure 3 for length saturation and manual recovery, Figure 4 for frequency prior and semantic routing, Figure 5 for full-suite controls, and Figure 6 for qualitative retention cases.Gives the paper a progressive order from interface and motivation to mechanism, controls, and evidence.
Cross-reference conventionsCaptions and nearby text now identify corresponding backbone references, equal-weight benchmark aggregation, checkpoint scope, and the distinction between token compression and measured decoding time.Prevents readers from conflating heterogeneous literature comparisons, macro averages, and system-level runtime evidence.
Supplementary organizationGrouped detailed reward definitions, routing features, replay details, evaluation protocol, checkpoint tables, MathVision trajectory, and case provenance under the current appendix structure.Keeps the main text focused while making each reported result traceable to a concrete protocol or case source.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.