acceptodds
Under review as a conference paper at ICLR 2027

Fused Multi-Latent Reasoning: Unified Multimodal Reasoning in Latent Space

Abstract

Vision-language reasoning often unfolds through intermediate steps that mix textual reasoning, visual evidence, and cross-modal grounding. Existing latent reasoning methods attempt to move such processes into latent space, but they typically rely on unimodal or weakly coupled supervision, leaving latent units poorly aligned with the multimodal reasoning steps they replace and prone to losing task-relevant visual evidence or semantic intent during compression. Our key insight is that high-dimensional vision-language embeddings can compress interleaved textual-visual reasoning sequences into dense representations that, after adaptation, serve as informative latent units for multimodal reasoning. In this work, we propose Fused Multi-Latent Reasoning (FMLR), a latent-space reasoning framework that learns latent units supervised by fused textual-visual step representations. By distilling step-level multimodal embeddings into latent units, FMLR enables the model to reason over implicit multimodal states. To further improve latent reasoning and control its length, we introduce Minimum-Sufficient Latent GRPO (MSL-GRPO), which combines answer correctness, threshold-based early sufficiency, and a per-unit latent cost to learn concise reasoning and explicit stopping decisions. Experiments on challenging vision-language reasoning benchmarks show that FMLR improves accuracy while producing substantially fewer output tokens than explicit-reasoning baselines. Code and datasets will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Reasoning, Latent Reasoning, Vision-Language Models, Cross-Modal Alignment, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_DRnmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how to improve multimodal reasoning while compressing explicit reasoning traces into short latent trajectories. It proposes FMLR, which supervises intermediate latent states using joint representations of reasoning text and corresponding visual evidence, together with MSL-GRPO to optimize latent reasoning, answer generation, and stopping through correctness, early-sufficiency, and length rewards. The authors also construct refined multimodal reasoning traces and corresponding embedding datasets. Evaluation includes eight vision-centric benchmarks, data-matched supervision and optimization ablations, an accuracy–output-length analysis, and experiments on two additional backbones. The reported results show accuracy gains for the complete training recipe and substantially shorter outputs than the Thinking and Refine CoT baselines.

Critical strengths
  1. This paper introduces a framework that supervises latent reasoning through joint embeddings of visual evidence and textual explanations, while incorporating reinforcement learning to optimize both reasoning and termination decisions. The approach strengthens visual reasoning without requiring extensive textual chains of thought. With the backbone and training data held constant, the full method achieves higher accuracy across all eight benchmarks, exceeding MSL with independent supervision by 0.5 to 7.7 percentage points and standard GRPO with fused supervision by 0.5 to 4.8 percentage points. On the four benchmarks reported in Table 3, it also outperforms Refine CoT by 1.6 to 5.0 percentage points and reduces the reported output length to approximately 9 to 10 tokens, compared with 156 to 217 tokens for Refine CoT. These results demonstrate its ability to deliver higher accuracy alongside substantially more concise outputs.

  2. This work develops reusable intermediate supervision by refining generated reasoning trajectories and aligning textual explanations with relevant image regions. The benefits of this refinement extend to explicit reasoning independently of the specific latent architecture used. Refine CoT achieves better results than Auto-CoT on seven of the eight benchmarks, with improvements of 9.4 percentage points on MME-RW and 9.9 percentage points . The resulting fused targets can also be reused across different backbones without requiring them to be embedded again. Using these targets, the full training method improves upon the corresponding base models on all eight benchmarks for InternVL3.5-8B and seven of the eight benchmarks for Qwen3-VL-4B. These findings highlight the practical value of constructing supervision that remains useful beyond the primary model.

Critical weaknesses
  1. The paper does not establish whether the early-sufficiency reward and learned stopping reduce latent computation without sacrificing accuracy. Table 2 reports accuracy gains, and Table 3 demonstrates compression relative to explicit CoT, but neither compares the actual latent-step counts of fused SFT, standard GRPO, and MSL-GRPO. Starting from the same SFT checkpoint and using the same latent/stop/answer policy, the authors should compare correctness-only reward, correctness plus a length penalty, and the full MSL reward, reporting both accuracy and average latent steps. Comparing learned stopping against fixed budgets of 2, 4, and 8 steps would further establish whether learning when to stop offers an advantage. These tests are needed to substantiate the specific benefits claimed for MSL-GRPO beyond the effectiveness of the complete training recipe.

  2. The comparison with prior latent reasoning methods leaves a consequential uncertainty about training conditions. Table 1 states that LVR-8B and Monet-8B use the same backbone as FMLR, but does not specify whether they receive the same refined training data or comparable training and inference budgets. This matters because Table 2 itself shows that data refinement substantially changes performance—for example, MME-RW rises from 42.0 to 51.4 under explicit CoT supervision. The authors should disclose the baseline configurations and, if these conditions are not already matched, compare FMLR with Monet using the same refined examples and comparable budgets. Although the internal ablations support FMLR’s effectiveness, this comparison is needed to determine how much its advantage over the closest prior method comes from the proposed method rather than differences in supervision or computation.

Questions for authors
  1. Could the authors clarify exactly how “Fused Latent + GRPO” differs from MSL-GRPO in Table 2? Does it optimize the same latent actions, stopping decisions, and answer tokens under the same rollout budget, and which reward terms does it use? If relevant ablations are already available, please provide accuracy and average latent-step counts with and without the early-sufficiency bonus and length penalty, together with any comparison of learned versus fixed-length stopping. These results would help establish whether the proposed reward and stopping mechanisms provide benefits beyond jointly optimizing latent reasoning and answer generation.

  2. Were the LVR-8B and Monet-8B baselines trained on the same refined Visual CoT examples and ZebraCoT subset as FMLR? Please clarify any differences in training budgets and inference settings, particularly the number of latent steps. If these conditions differ, are comparisons under matched conditions already available? This clarification would affect my confidence that the improvements over prior methods come from the proposed method rather than differences in training data or computation.

4 comments
AuthorsOpenReview ↗

1. W1 / Q1 — Correctness-only GRPO, MSL rewards, and learned stopping

Thank you for the suggestion. We now report the correctness-only baseline, all reward combinations, actual latent counts, and fixed-budget inference results.

What changes from GRPO to MSL-GRPO. The original “Fused Latent + GRPO” row is correctness-only C, now labeled in Table 2. C, C+L, C+S, and Full start from the same fused-SFT checkpoint. All optimize latent actions, stop/continue decisions, and answer tokens, using four trajectories per prompt and a shared trajectory-level group advantage. MSL-GRPO adds the early-sufficiency bonus and per-latent length cost to this policy.

The reward objectives are C: R = C; C+L: R = C − 0.1O/8; C+S: R = C(1 + 0.3S); Full: R = C(1 + 0.3S) − 0.1O/8. C is final-answer correctness; O is generated latent count. S = max(1 − k⋆/8, 0), with k⋆ the first prefix reaching correct-option probability 0.7 and k⋆ = 9 if there is no crossing. Thus, C+L removes sufficiency, C+S removes length cost, and C retains correctness alone.

The update uses standard GRPO group-relative advantages: each trajectory's reward is centered and scaled within its four-rollout group. Its latent actions, sampled stop decisions, and answer tokens all share that advantage. Equal rewards give zero policy advantage. At inference, the learned policy selects stop/continue directly from its logits.

SFT and every reward variant across all benchmarks. The SFT column is from Table 2; the reward columns are from Table 3. Scores are accuracy (%) ↑.

BenchmarkNFused SFTCC+LC+SFull
MMVP30078.378.379.6780.3380.3
HR-4K80073.977.978.5077.7578.8
HR-8K80071.867.172.6375.0071.9
AI2D3,08879.282.383.9779.7083.3
MME-RW Lite1,91955.455.861.1358.0058.4
V*19183.283.881.1585.3484.3
TextVQA5,00078.080.180.9679.8681.9
VisualPuzzles1,16833.336.438.5333.3937.2

Table 3 summarizes the reward variants:

MetricCC+LC+SFull
8-benchmark macro (%) ↑70.21372.06671.17172.013
Mean latents ↓2.0001.7001.5501.000

The fused-SFT native-generation analysis records 2.000 latent states per question over 13,266 questions. Full improves over fused SFT on all eight benchmark scores, including HR-4K 73.9→78.8, MME-RW 55.4→58.4, and VisualPuzzles 33.3→37.2. Against correctness-only C, Full gains 1.800 macro points while halving latent length. C+L reaches 72.066% with 1.700 latents and C+S 71.171% with 1.550; Full retains competitive macro accuracy with the shortest mean trajectory.

Does learned stopping help at inference? We keep the final RL checkpoint, answer decoder, and 700 questions fixed and compare learned stopping with exactly 2, 4, or 8 latent actions. The panel has 100 questions each from MMVP, HR-4K, HR-8K, MME-RW Lite, V*, VisualPuzzles, and MMStar. Table 5 reports:

StoppingCorrect/NAccuracy % ↑Latents ↓Mean s ↓
Learned493/70070.4291.02.706
Fixed 2481/70068.7142.02.710
Fixed 4479/70068.4294.02.822
Fixed 8479/70068.4298.03.033

Learned stopping answers 12–14 more questions correctly and reduces mean latency by 10.8% compared with eight fixed steps. The model switches to answer generation after its selected stop decision; fixed-length conditions switch after the prescribed number of latent transitions.

Manuscript changes. Table 2, p. 8 identifies C and the SFT baseline; §4.3, Tables 3 and 5, pp. 9–10 adds reward/length and stopping comparisons. §3.4, pp. 5–7 and Appendix E, Table 13, p. 33 specify the action blocks, advantage, and rewards. The revised manuscript is provided in the Supplementary Material with changes highlighted in blue.

AuthorsOpenReview ↗

2. W2 / Q2 — Shared supervision and training/inference budgets for LVR and Monet

Thank you for emphasizing the effect of data refinement. The LVR-8B and Monet-8B entries are our reimplementations on the same Qwen3-VL-8B backbone, refined Visual CoT examples, and ZebraCoT subset as FMLR. We now state this configuration explicitly and disclose their training and inference budgets.

Same refined supervision. Table 2 shows the value of refinement: explicit Auto-CoT and Refine CoT achieve 42.0% and 51.4% on MME-RW, respectively. All three latent methods use the refined data. Stage 1 contains 403,593 examples; Stage 2 uses the ZebraCoT subset.

Complete eight-benchmark results from Table 1. Accuracy (%) ↑:

BenchmarkNLVR-8B ↑Monet-8B ↑FMLR-8B ↑
MMVP30077.778.380.3
HR-4K80070.076.278.8
HR-8K80063.870.071.9
AI2D3,08881.782.383.3
MME-RW Lite1,91950.353.658.4
V*19182.783.284.3
TextVQA5,00080.680.181.9
VisualPuzzles1,16834.436.937.2

FMLR improves over both reimplemented methods on all eight benchmarks. Against Monet, the gains include 2.0 points on MMVP, 2.6 on HR-4K, and 4.8 on MME-RW; the MME-RW gain over LVR is 8.1 points.

Inference budgets. LVR-8B and Monet-8B each execute eight latent steps. FMLR learns to stop within one to eight steps. Table 4 uses LVR-7B for the timing comparison. Table 5 compares learned and fixed 2/4/8-step inference within FMLR.

SFT budget and measured time. All methods use 403,593 examples. FMLR completes one epoch in 1,577 updates; LVR uses the same sample count; Monet follows its three-stage schedule. Table 16 records wall-clock hours on NVIDIA A800 80GB GPUs:

ModelStage 1 h ↓Stage 2 h ↓Stage 3 h ↓Total h ↓
FMLR-8B34.83——34.83
LVR-8B42.1——42.1
Monet-8B104.566.347.5218.3

RL exposure and measured time. The common one-epoch budget contains 5,872 prompts and four trajectories per prompt. Batch sizes lead to different update counts, which Table 17 reports explicitly:

ModelPromptsRollouts/promptUpdatesRL h ↓
FMLR-8B5,87247343.12
LVR-8B5,872473413.0
Monet-8B (VLPO)5,87249218.3

SFT and RL timing exclude offline target precomputation, evaluation, and judging. FMLR takes 34.83 h for SFT and 3.12 h for RL under these configurations, with the highest accuracy across all eight benchmarks.

Manuscript changes. §4.1, p. 7 specifies the common backbone, refined examples, ZebraCoT subset, and eight-step baseline inference. Table 1, p. 8 provides all benchmark scores; Appendix G, Tables 16–17, p. 34 gives stage-wise SFT and RL budgets. The revised manuscript is provided in the Supplementary Material with changes highlighted in blue.

Thank you for the additional experiments and clarifications. I still have two questions.

1. Does learned stopping actually achieve adaptive reasoning?

Both the Full reward configuration and the learned stopping experiment report an average of 1.000 latent steps. Does this mean the model always stops after the first step? If so, it seems to behave like a fixed one-step strategy rather than an adaptive stopping policy. Could the authors provide the distribution of stopping steps and results for a Fixed-1 baseline?

2. What is the main benefit of the learned stopping mechanism?

C+L achieves slightly higher average accuracy than Full (72.066% vs. 72.013%), while Full uses fewer latent steps. Also, learned stopping has nearly the same latency as Fixed-2 (2.706s vs. 2.710s). Could the authors clarify whether the main benefit is accuracy or efficiency? A comparison under matched latent-step budgets would help establish the actual contribution of learned stopping.

Thank you for these follow-up questions. Our objective is straightforward: help the model answer correctly as early as possible, while avoiding unnecessary latent computation.

1. Learned stopping and effective one-step reasoning

MSL-GRPO jointly rewards answer correctness and early sufficiency while discouraging unnecessary latent steps. The model can use between one and eight steps and learns when to stop. On the reported 700-question panel, it consistently chooses to answer after its first latent step: 700 questions use one step (100%), and none use two to eight steps.

The reported comparisons show that this is an effective operating point. Learned stopping achieves 70.429% accuracy, exceeding Fixed-2 by 1.71 percentage points and both Fixed-4 and Fixed-8 by 2.00 percentage points. What matters is how effectively the model uses its latent reasoning, rather than the number of steps alone.

We further compare the reward configurations when all conditions execute one latent step on the same 700 questions:

Table 1. Accuracy of MSL-GRPO reward configurations at a matched one-step budget on 700 questions.

Reward configurationStopping ruleCorrect / TotalAccuracy
CFixed-1469/70067.000%
C+LFixed-1486/70069.429%
C+SFixed-1482/70068.857%
FullLearned stopping (one step on every question)493/70070.429%

Full achieves the highest accuracy among these four configurations at the same one-step budget, outperforming C, C+L, and C+S by 3.43, 1.00, and 1.57 percentage points, respectively. This shows that the complete reward design produces a more effective one-step answering policy.

The comparison with non-latent reasoning provides complementary evidence. In Table 2, Full outperforms Non-latent CoT + GRPO on all eight benchmarks, increasing macro accuracy from 67.538% to 72.013%, an improvement of 4.475 percentage points.

Together, these comparisons highlight the value of learning to use latent computation effectively. The policy learns its stopping decision within the available budget; on this panel, that learned decision is consistently one step. The resulting policy achieves strong performance with a compact trajectory and benefits from the complete reward design even when the compared configurations use the same number of latent steps.

2. Accuracy and efficiency

Table 1 shows how effectively each reward configuration uses a single latent step. With all three MSL-GRPO reward components, Full learns to stop at the earliest permitted step on all 700 questions. It achieves the highest accuracy among the four configurations at this matched budget (70.429%), exceeding C, C+L, and C+S by 3.43, 1.00, and 1.57 percentage points, respectively.

Across the eight benchmarks, Full maintains nearly the same macro accuracy as C+L (72.013% versus 72.066%) while using 41.2% fewer latent steps on average (1.00 versus 1.70). On the 700-question panel, learned stopping and Fixed-2 have nearly identical mean latency (2.706 s versus 2.710 s), while learned stopping achieves 1.71 percentage points higher accuracy with half as many latent steps. Compared with Fixed-8, it improves accuracy by 2.00 percentage points and reduces mean latency by 10.8%.

These 8B results motivate exploring compressed reasoning at larger scales. For a dense 100B decoder, saving one decoding step per answer would avoid roughly 200 billion floating-point operations in matrix multiplications alone. This estimate illustrates the potential value of preserving accuracy with shorter reasoning sequences across many queries. This makes preserving accuracy while reducing the total number of decoding steps a promising direction for future work. Pope et al., Efficiently Scaling Transformer Inference, MLSys 2023

Reviewer_aFY2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Fused Multi-Latent Reasoning, which supervises continuous latent states using embeddings that jointly encode reasoning statements and visual evidence. Stage 1 combines hidden-state alignment with answer supervision. Stage 2 introduces Minimum-Sufficient Latent GRPO, jointly optimizing latent actions, stopping decisions, and answer tokens through correctness, early sufficiency, and latent-length rewards. Experiments on eight vision-centric benchmarks show accuracy improvements and short output sequences, with additional evaluations on InternVL3.5-8B and Qwen3-VL-4B.

Critical strengths
  1. Meaningful gains with relevant controls. FMLR consistently improves over the base model across the evaluated benchmarks. Comparisons with LVR and Monet using the same backbone, matched training data, and evaluations on additional backbones strengthen the empirical evidence.

  2. A concrete approach to multimodal latent supervision. Pairing reasoning statements with visual evidence provides explicit targets for latent alignment. Fused supervision performs better than separate supervision under MSL-GRPO, suggesting complementary benefits from representation and optimization.

Critical weaknesses
  1. Efficiency and adaptive stopping are insufficiently validated. The token-counting convention is unclear, and output length alone cannot establish inference savings. Latency measurements and latent-length statistics are needed to determine whether learned stopping reduces computation while preserving accuracy.

  2. The experiments do not sufficiently isolate the proposed mechanisms. The contributions of the sufficiency bonus, length penalty, and learned stopping remain unclear. Similarity and noise-perturbation analyses do not establish semantic grounding and examine trajectories before reinforcement learning rather than the final model. Component ablations and targeted controls would strengthen the explanation of the gains.

Questions for authors
  1. What exactly do Table 3's token counts include? Please report latency, average latent-action count, and the fraction reaching the maximum budget under matched inference settings.

  2. How do accuracy and latent length change when independently removing the sufficiency bonus or length penalty, or replacing learned stopping with fixed budgets?

  3. For the final RL model, what happens when latent states are exchanged between examples or replaced with position-wise means while keeping computation length fixed?

  4. How are reference steps mapped to latent slots, what enters each slot during Stage 1, and how are targets normalized or projected across backbones? Specify training hyperparameters and configurations for the reimplemented baselines.

5 comments
AuthorsOpenReview ↗

1. W1 / Q1 — Token accounting, latency, and adaptive stopping

Thank you for asking us to measure computation directly and clarify what is counted. We now present accuracy, output tokens, latency, and actual latent-action counts together, supplemented by generation-limit and stopping statistics.

Token convention. We count generated output IDs and continuous latent actions separately. Output IDs include answer tokens and output markers appended after the prompt. Extracted-answer length is retokenized without extra special tokens. Latent count is the number of continuous actions executed before answer generation.

Matched timing protocol. Six models run sequentially on eight NVIDIA A800 80GB GPUs, with the same 2,091 questions, preprocessing, 1,024–10,035,200-pixel image budget, 16,384-token input limit, BF16, SDPA, batch size one per device, and greedy decoding with one beam and repetition penalty one. Timing includes image reading, preprocessing, transfer, synchronized generation, and decoding, after warmup. Loading, warmup, scoring, and logging are excluded.

Per-benchmark accuracy and length, plus full latency statistics. Table 4 below reports accuracy (%) ↑ / mean output tokens ↓ in each benchmark cell. Latency is in seconds. The panel contains MMVP 300, HR-4K 800, HR-8K 800, and V* 191 questions.

ModelMMVPHR-4KHR-8KV*Mean s ↓Median s ↓P95 s ↓Latents ↓
Qwen3-VL-8B Instruct76.7/91.675.6/168.069.1/169.980.6/4.18.9207.50343.841—
Qwen3-VL-8B Thinking76.3/302.268.9/389.063.5/422.977.0/182.816.66611.09648.380—
Visual CoT SFT77.3/7.672.3/31.766.4/25.883.8/2.26.3885.59914.788—
Refine CoT SFT75.3/217.475.6/197.470.2/208.382.7/156.46.9735.58515.691—
LVR-7B71.7/24.970.8/29.163.0/29.981.7/20.05.9267.1968.1998.0
FMLR-8B80.3/9.978.8/9.071.9/9.084.3/9.05.0116.1647.0491.0

FMLR achieves the highest accuracy on each benchmark, with the lowest mean latency (5.011 s) and P95 (7.049 s). Its mean latency is 15.4% below LVR-7B, alongside one latent action versus eight.

Maximum-budget and truncation checks. FMLR uses one latent action on all 2,091 questions; 0/2,091 (0%) reach its eight-step maximum. Table 14 records output limits and truncation:

ModelGeneration capTruncated/N
Qwen3-VL-8B Instruct1,024110/2,091
Qwen3-VL-8B Thinking1,024223/2,091
Visual CoT SFT1,02444/2,091
Refine CoT SFT1,02474/2,091
LVR-7B1,0240/2,091
FMLR-8B640/2,091

FMLR's cap applies to answer tokens; the other caps apply to generated IDs. Scoring uses the unedited output.

Table 5 compares stopping rules using the same FMLR checkpoint and 700 questions. Learned stopping obtains 70.429% with one latent action and 2.706 s; fixed 2/4/8 steps obtain 68.714/68.429/68.429% at 2.710/2.822/3.033 s. Learned stopping achieves higher accuracy with 10.8% lower latency than the eight-step configuration.

Manuscript changes. §4.3, Tables 4–5, p. 10 presents efficiency and stopping; Appendix E, Tables 13–14, p. 33 documents execution and counting. The revised manuscript is provided in the Supplementary Material with changes highlighted in blue.

AuthorsOpenReview ↗

2. W2 / Q2–Q3 — Rewards, stopping, and final-model latent content

Thank you for the suggestions. We added reward ablations, fixed-budget stopping comparisons, and input-latent interventions in the final RL model.

Reward components. C, C+L, C+S, and Full start from the same fused-SFT checkpoint and optimize the same latent/stop/answer action blocks with four trajectories per prompt. Their objectives are C: R = C; C+L: R = C − 0.1O/8; C+S: R = C(1 + 0.3S); Full: R = C(1 + 0.3S) − 0.1O/8. C is final-answer correctness, O counts generated latent actions, and S = max(1 − k⋆/8, 0). The first prefix reaching correct-option probability 0.7 defines k⋆; if there is no crossing, k⋆ = 9. Full→C+L removes sufficiency, and Full→C+S removes the length cost. Table 3 gives all eight benchmark accuracies (%) ↑:

BenchmarkNCC+LC+SFull
MMVP30078.379.6780.3380.3
HR-4K80077.978.5077.7578.8
HR-8K80067.172.6375.0071.9
AI2D3,08882.383.9779.7083.3
MME-RW Lite1,91955.861.1358.0058.4
V*19183.881.1585.3484.3
TextVQA5,00080.180.9679.8681.9
VisualPuzzles1,16836.438.5333.3937.2
8-benchmark macro (%) ↑13,26670.21372.06671.17172.013
Mean latents ↓—2.0001.7001.5501.000

Full gains 1.800 macro points over C while halving mean latent length. C+L and C+S also reduce latent length, with complementary benchmark strengths: C+L reaches 83.97% on AI2D and 38.53% on VisualPuzzles, while C+S reaches 75.00% on HR-8K and 85.34% on V*. Full combines competitive accuracy with the shortest mean trajectory.

Learned versus fixed stopping. On the same final checkpoint and 700 questions, the answer decoder is held fixed and the inference rule selects learned stopping or exactly 2/4/8 latent actions. Table 5 reports:

StoppingCorrect/NAccuracy % ↑Latents ↓Mean s ↓
Learned493/70070.4291.02.706
Fixed 2481/70068.7142.02.710
Fixed 4479/70068.4294.02.822
Fixed 8479/70068.4298.03.033

Learned stopping answers 12–14 more questions correctly and reduces mean latency by 10.8% versus eight fixed steps. Ground-truth answers determine training rewards; inference selects stopping from the learned policy's logits.

Targeted interventions in the final RL model. We evaluate the requested cross-example and position-wise mean replacements after RL, using one latent step throughout. We replace the next-step input vector while retaining the recipient's image, question, and model parameters. The panel contains 100 questions each from MMVP, HR-4K, HR-8K, MME-RW Lite, V*, VisualPuzzles, and MMStar. Table 9 reports:

Input latentNAccuracy % ↑Latents
Original70070.4291
Cross-example replacement70066.0001
Position-wise mean replacement70065.0001

Replacing the mixed text-visual input latent lowers accuracy by 4.429 and 5.429 points at the same computation length, showing that latent content contributes to final-answer accuracy. Table 10 holds step text fixed while exchanging crops in the fused encoder: the resulting mean relative L2 is 0.5716 and cosine similarity is 0.8261 across 95 step pairs.

Manuscript changes. §4.3, Table 3, p. 9 covers rewards; Table 5, p. 10 covers stopping; Appendix A.5–A.6, Tables 9–10, p. 18 covers content interventions and crop encoding. The revised manuscript is provided in the Supplementary Material with changes highlighted in blue.

AuthorsOpenReview ↗

Q4 — Slot construction, alignment targets, and implementation settings

Thank you for requesting the implementation details. We expanded the alignment specification and configuration tables so that the step-to-slot mapping, Stage 1 inputs, target dimensions, and baseline settings are explicit.

Reference steps and slot inputs. Each retained nonempty reference step supplies one target in original order. Its reasoning statement and paired visual evidence are jointly encoded into a fused target. The student sequence contains a learned placeholder at each corresponding latent slot. Stage 1 uses one causal forward pass: the hidden state immediately preceding each placeholder is supervised against the paired target through the alignment loss. The target enters that loss; the placeholder supplies the slot input. At inference, each predicted latent is fed into the next step.

Alignment and answer objectives. For example i with N_i targets and dimension d, alignment is the sum of squared hidden/target differences divided by N_i d. The batch objective averages this example-level term, multiplies it by 0.1, and adds cross-entropy on answer tokens. Intermediate reasoning text and latent placeholders are masked from answer supervision.

Dimensions and normalization. The 8B alignment uses 4,096-dimensional targets and direct MSE. The 4B hidden state is 2,560-dimensional and is mapped to 4,096 dimensions through LayerNorm and an MLP. The target loader adds no L2 normalization. The target resource is reused across backbones, with the student-side projection accommodating the dimension change.

FMLR settings (Table 13).

SettingConfiguration
SFT alignment / learning rate0.1 / 1e-5
SFT batch / accumulation / duration4 per device / 8 / 1 epoch
SFT precision / frozen modulesBF16 / visual encoder and multimodal projector
RL initialization / dataSame fused-SFT checkpoint / ZebraCoT subset
RL adaptersLanguage attention and MLP; LoRA rank 16, scaling 32
RL optimizerAdamW; learning rate 5e-6; weight decay 0.01
Schedule / clippingWarmup 0.03; gradient clip 1.0; PPO clip 0.2
KL / latent noise0.01 / standard deviation 0.5
Full rewardThreshold 0.7; sufficiency weight 0.3; length weight 0.1
Rollouts / stopping budget4 per prompt / 1–8 latent actions

Latent, stop, and answer actions share a trajectory-level group advantage. The frozen Stage 1 model is the reference. At inference, latent actions use their means, stopping selects the larger logit, and answer decoding is greedy.

Reimplemented latent baselines. LVR-8B and Monet-8B use the same Qwen3-VL-8B backbone, refined Visual CoT examples, and ZebraCoT subset as FMLR. Both use eight latent steps at inference. Their SFT data contain 403,593 examples. Measured SFT times are 42.1 h for LVR and 218.3 h across Monet's three stages, compared with 34.83 h for FMLR. Their common one-epoch RL exposure is 5,872 prompts with four trajectories each; update counts are 734 for LVR/FMLR and 92 for Monet, reflecting their batch sizes. RL times are 13.0, 3.12, and 18.3 h, respectively. Table 1 reports LVR-8B and Monet-8B; Table 4 uses LVR-7B.

Manuscript changes. §3.3, pp. 4–5 and Eq. (1) specify alignment; §3.4, pp. 5–7 specifies RL; Appendix E, Table 13, p. 33 lists settings. §4.1, p. 7 and Appendix G, Tables 16–17, p. 34 disclose the baseline data, stages, and budgets. The revised manuscript is provided in the Supplementary Material with changes highlighted in blue.

Reviewer_TA1MOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses two challenges in multimodal reasoning: the length of explicit image–text reasoning traces, and the difficulty of ensuring that a latent reasoning unit captures both semantic intent and its paired visual evidence. FMLR aligns latent states with fused text–visual representations of individual reasoning steps. It then uses MSL-GRPO to optimize answer correctness, reasoning length, and stopping decisions.

Critical strengths
  1. The paper is clearly written and easy to follow.
  2. The results show higher reported accuracy than the authors’ same-backbone reimplementations of LVR and Monet across the eight evaluated benchmarks.
Critical weaknesses
  1. The proposed mechanism is not directly validated. The fused-versus-separate ablation and random latent perturbations do not establish that the latent units retain the specific visual evidence paired with each reasoning step.
Questions for authors
  1. How do the latent representations and answer accuracy change if the paired image crops are mismatched or replaced while the reasoning text is held fixed?
  2. What is the accuracy–latency trade-off under matched hardware, image resolution, and decoding settings? Please include the average number of latent steps.
  3. What are the separate effects of the sufficiency bonus, latent-length penalty, and learned stopping policy? How does the method compare with fixed latent-step budgets?
  4. How were image and question overlaps between the training data and evaluation benchmarks identified and removed?
4 comments
AuthorsOpenReview ↗

1. W1 / Q1 — Visual evidence and answer accuracy

Thank you for the suggestion. We added input-latent replacements in the final RL model and a fixed-text crop-encoding control to examine the contribution of latent content and paired visual evidence.

Effect on final answers. We evaluate the final RL model on 700 questions: 100 each from MMVP, HRBench-4K, HRBench-8K, MME-RealWorld Lite, V*, VisualPuzzles, and MMStar. We replace the vector supplied to the next latent input position with either another example's latent or a position-wise mean. The recipient image, question, model parameters, and one-step computation budget remain fixed. Table 9 reports:

Input latentNAccuracy % ↑Latents
Original70070.4291
Cross-example replacement70066.0001
Position-wise mean replacement70065.0001

Cross-example replacement lowers accuracy by 4.429 percentage points, and position-wise mean replacement lowers it by 5.429 points. Both replacements affect the mixed text-visual input latent at the same one-step budget, showing that its content contributes to answer accuracy.

Visual contribution to the fused representation. To examine the paired visual component, we hold the reference step text, canvas size, and prompt-token count fixed while exchanging the crop. The control covers 95 steps from 32 Flickr30k training-reference images. Matched/mismatched encodings and six identical-input repeats require 196 calls. Table 10 reports:

Encoding controlPairsMean relative L2Mean cosine
Matched/mismatched crop950.57160.8261
Identical-input repeat60.0206—

Changing the crop under fixed step text changes the fused representation: mean relative L2 is 0.5716 for crop exchange and 0.0206 for identical-input repeats. Crop-exchange distances exceed repeat distances in all six repeat-controlled steps. Repeat distances range from 0.0151 to 0.0245, above the prespecified stability tolerance of 0.001.

Manuscript changes. Appendix A.5, Table 9, p. 18 adds the final-model input-latent results; Appendix A.6, Table 10, p. 18 adds the fixed-text crop control. The revised manuscript is provided in the Supplementary Material with changes highlighted in blue.

AuthorsOpenReview ↗

2. Q2 — Accuracy, output length, latency, and latent computation

Thank you for requesting a direct accuracy–latency comparison. Table 4 now combines per-benchmark accuracy, output-token counts, measured latency, and latent-step counts.

Matched execution. The six models run sequentially on eight NVIDIA A800 80GB GPUs on the same 2,091 questions: MMVP 300, HRBench-4K 800, HRBench-8K 800, and V* 191. They share preprocessing, a 1,024–10,035,200-pixel image budget, a 16,384-token input limit, BF16, SDPA, batch size one per device, and greedy decoding with one beam and repetition penalty one. Timing starts with image reading and includes preprocessing, transfer, synchronized generation, and decoded output. Model loading, warmup, scoring, and logging are excluded.

Complete four-benchmark comparison. Each benchmark cell below is accuracy (%) ↑ / mean output tokens ↓; latency columns are seconds over the 2,091-question panel. Output tokens and continuous latent actions are counted separately.

ModelMMVPHR-4KHR-8KV*Mean s ↓Median s ↓P95 s ↓Latents ↓
Qwen3-VL-8B Instruct76.7/91.675.6/168.069.1/169.980.6/4.18.9207.50343.841—
Qwen3-VL-8B Thinking76.3/302.268.9/389.063.5/422.977.0/182.816.66611.09648.380—
Visual CoT SFT77.3/7.672.3/31.766.4/25.883.8/2.26.3885.59914.788—
Refine CoT SFT75.3/217.475.6/197.470.2/208.382.7/156.46.9735.58515.691—
LVR-7B71.7/24.970.8/29.163.0/29.981.7/20.05.9267.1968.1998.0
FMLR-8B80.3/9.978.8/9.071.9/9.084.3/9.05.0116.1647.0491.0

FMLR achieves the highest accuracy on all four benchmarks and the lowest mean and P95 latency among the evaluated configurations. Its mean latency is 5.011 s versus 5.926 s for LVR-7B, a 15.4% reduction. Its P95 is 7.049 s versus 8.199 s. FMLR uses 9.9 output tokens on MMVP and 9.0 on the other three benchmarks, alongside one latent action; LVR-7B uses eight latent steps.

Actual stopping behavior. All 2,091 FMLR questions use one latent action, and 0/2,091 reach its maximum budget of eight. FMLR completes all outputs within its generation limit. Appendix E records the generation caps and truncation counts: the five comparison configurations use a 1,024 generated-ID cap; FMLR uses a 64-answer-token cap. Table 5 compares learned stopping with fixed lengths using the same checkpoint.

Manuscript changes. Table 4 and the efficiency analysis are in §4.3, p. 10; execution settings and generation limits are in Appendix E, Tables 13–14, p. 33. The revised manuscript is provided in the Supplementary Material with changes highlighted in blue.

AuthorsOpenReview ↗

3. Q3 — Reward components and learned stopping

Thank you for the suggestion. We added reward-component ablations and a fixed-budget inference comparison using the same final checkpoint.

Reward definitions and common initialization. All four reward variants start from the same fused-SFT checkpoint and optimize latent actions, stop/continue decisions, and answer tokens. Four trajectories are sampled per prompt, and the three action blocks share the trajectory's group-relative advantage. C rewards correctness; C+L adds the latent-length cost; C+S adds early sufficiency; Full combines all three.

The objectives are C: R = C; C+L: R = C − 0.1O/8; C+S: R = C(1 + 0.3S); Full: R = C(1 + 0.3S) − 0.1O/8. Here C is final-answer correctness and O is the generated latent count. S = max(1 − k⋆/8, 0), where k⋆ is the first prefix whose correct-option probability reaches 0.7; k⋆ = 9 if there is no crossing. The sufficiency bonus is gated by correctness, and the length cost applies to every latent action. Full→C+L removes sufficiency; Full→C+S removes the length cost.

All eight benchmarks. Table 3 reports accuracy (%) ↑, equal-weight benchmark macro accuracy ↑, and average latent count ↓:

BenchmarkNCC+LC+SFull
MMVP30078.379.6780.3380.3
HR-4K80077.978.5077.7578.8
HR-8K80067.172.6375.0071.9
AI2D3,08882.383.9779.7083.3
MME-RW Lite1,91955.861.1358.0058.4
V*19183.881.1585.3484.3
TextVQA5,00080.180.9679.8681.9
VisualPuzzles1,16836.438.5333.3937.2
8-benchmark macro (%) ↑13,26670.21372.06671.17172.013
Mean latents ↓—2.0001.7001.5501.000

Full improves macro accuracy by 1.800 points over C while reducing mean latent length from 2.000 to 1.000. C+L reaches 72.066% with 1.700 latents, and C+S reaches 71.171% with 1.550. Full retains competitive accuracy relative to C+L with fewer latent actions. The detailed rows expose the task-specific contributions: C+L reaches 83.97% on AI2D and 38.53% on VisualPuzzles; C+S reaches 75.00% on HR-8K and 85.34% on V*; Full reaches 80.3% on MMVP, 78.8% on HR-4K, and 81.9% on TextVQA with the shortest mean trajectory.

Stopping policy at fixed model weights. We then keep the final RL checkpoint, answer decoder, and 700 questions fixed, changing only whether the policy stops naturally or generates exactly 2, 4, or 8 latent actions before answering. The panel contains 100 questions from each of seven benchmarks. Table 5 reports:

StoppingCorrect/NAccuracy % ↑Latents ↓Mean s ↓
Learned493/70070.4291.02.706
Fixed 2481/70068.7142.02.710
Fixed 4479/70068.4294.02.822
Fixed 8479/70068.4298.03.033

Learned stopping answers 12–14 more questions correctly and reduces mean latency by 10.8% relative to eight fixed steps. The sufficiency threshold defines the training bonus; at inference, the policy selects stop/continue directly from its logits.

Manuscript changes. Reward ablations appear in §4.3, Table 3, p. 9; learned/fixed stopping appears in Table 5, p. 10. Reward and optimization definitions are in §3.4, pp. 5–7, and Appendix E, Table 13, p. 33. The revised manuscript is provided in the Supplementary Material with changes highlighted in blue.

AuthorsOpenReview ↗

4. Q4 — Image and question overlap identification and removal

Thank you for asking about image and question overlap. We report the checking procedure and records for 403,593 Stage 1 Visual CoT records and the 5,872-record Stage 2 ZebraCoT subset. The checks cover all 13,266 questions from the eight main benchmarks, together with MMStar used in the representation analysis.

Identification procedure. Dataset-qualified photo identifiers, file-byte hashes, and decoded RGB hashes identify reused photos, including copies with different file encodings. Candidate image variants receive visual verification. We check questions, options, answers, and paraphrases together with the matched photo.

Benchmark-level results. Table 15 summarizes verified photo/variant matches by the number of distinct evaluation questions affected. The seven-benchmark zero-match row is expanded below into its constituent benchmarks:

Evaluation setNStage 1 photo/variant hitsStage 2 photo/variant hits
MMVP30000
HR-4K80000
HR-8K80000
AI2D3,08800
MME-RW Lite1,91900
V*19100
TextVQA5,000810
VisualPuzzles1,16800
Main-eight total13,266810
MMStar parent set1,50091

Stage 2 has zero verified photo overlaps with the eight main benchmarks. For Stage 1, the 81 TextVQA questions comprise 69 questions sharing 43 byte-identical photos and 12 questions using eight visually verified variants. MMStar contributes nine Stage 1 matches and one Stage 2 match in its 1,500-question parent set.

Removal. We removed the identified images and associated QA examples from the current training manifests, including the ZebraCoT–MMStar item that uses the same photo and asks about the same fact. Removal covers the linked QA examples and their reformulations.

Manuscript changes. Appendix F, Table 15, p. 34 records the source sizes, matching procedure, affected-question counts, and removal from current manifests. The revised manuscript is provided in the Supplementary Material with changes highlighted in blue.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Review Discussion and Author Responses for the Area Chair

Dear Area Chair,

Thank you for considering our submission. We summarize below the main points raised during review and how we addressed them. The reviewers recognized the value of combining visual evidence with reasoning text to supervise latent representations, and highlighted FMLR’s ability to produce accurate, concise answers.

Reviewer TA1M asked how visual evidence affects reasoning and how we checked for overlap between training and evaluation data. We showed that changing the image crop while holding the reasoning text fixed changes the fused representation. In a separate experiment, replacing the final model’s latent with one from another example lowered answer accuracy by 4.43 percentage points. We also documented our overlap checks.

In response to Reviewer aFY2, we compared latency across six models, reported output lengths and latent-step counts, and evaluated the individual reward components. Reviewer aFY2 confirmed that our response addressed their concerns and maintained their positive recommendation for acceptance.

Reviewer DRnm acknowledged the additional experiments and clarifications, then asked what the model learns about when to stop and how early stopping affects accuracy and efficiency. FMLR chose to answer after one latent step on all 700 questions. At that same one-step budget, Full achieved 70.43% accuracy versus 69.43% for the strongest reward ablation. Across the eight benchmarks, Full averaged 1.00 latent step versus 1.55–2.00 for the ablations. On the 700-question panel, its latency was similar to Fixed-2 and 10.8% lower than Fixed-8.

Future work could explore latent reasoning in models with hundreds of billions of parameters, where generating even one fewer token per answer could yield substantial compute savings across many queries.

The revised manuscript is available in the Supplementary Material, with changes highlighted in blue. Thank you for your time and consideration.

Best regards,
All authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.