Answers right, edits ignored: ReaLVR adds evidence supervision to latent visual reasoning, 63.7% avg

Rethinking Latent Visual Reasoning: Grounding Latent Reasoning in Visual Evidence

Xi Xiao, Tianchen Zhao, Youngeun Kim, Zhuowei Li, Linghan Xu, Jiaye Wu, Zheng Zhang, Xiang Xu, Xuanbai Chen, Farhan Tejani, Jakub Zablocki, Julia Xu, Yifan Xing

cs.CV, cs.CL, cs.LG

2026-09-28

Latent visual tokens barely react to answer-changing edits; ReaLVR adds evidence supervision to GRPO trajectories: 63.7% five-task avg on Qwen2.5-VL-7B, gains hold at 235B.

What problem this solves

Latent visual reasoning (LVR) lets a multimodal LLM do intermediate computation in continuous latent tokens instead of writing out chain-of-thought. It saves output tokens and suits spatial or fine-grained visual judgments that are hard to verbalize. The cost is opacity: a textual rationale can be read and audited, a latent span can only be judged by whether the final answer is right.

The diagnostics here are blunt. Across four edit types (color change, object removal, shape swap, left-right swap), each with 512 original-edited pairs, the correct answer flips in 81.45%–86.33% of pairs while vanilla LVR changes its prediction in only 5.66%–13.09%. The mean-pooled latent trajectory barely moves either: cosine distance between original and edited inputs stays below 0.0005. A right answer guarantees nothing about what the latent tokens preserved.

The cause sits in training. Stage 1 supplies visual targets under teacher forcing; Stage 2 GRPO scores only the generated text, replaying sampled latent vectors as fixed context, so no gradient reaches the process that produced them. An outcome reward says a rollout was right; it says nothing about which latent position should hold which evidence. The paper names this the latent evidence-credit gap.

Method

ReaLVR adds one visual-evidence loss on top of Stage-2 GRPO and supervises a differentiable latent trajectory the model just generated itself. Architecture and inference stay untouched. Three choices:

• Supervise the model's own trajectory. The current model regenerates K = 8 latent tokens by recursively feeding back its own hidden states, with gradients kept through the whole recurrence, so the evidence loss updates the process that produces latents. Supervising saved rollout latents instead costs 1.8 points, the largest ablation drop. The principle is on-policy distillation: at inference the model generates its own prefix, so supervision must land on that prefix.
• What to preserve, by visual contrast. An ROI annotation is pooled into a positive prototype; 16 prototypes pooled from other examples in the batch serve as negatives. Each latent state must beat its most similar negative by a cosine margin of 0.5; without an ROI the target degrades to the whole image. Dropping the negatives costs 1.6 points, since plain alignment pulls latents toward generic image content.
• Where to supervise, by answer contrast. The correct answer and distinct wrong answers sampled by the behavior policy are teacher-forced in separate branches after the same regenerated span. Attention to each latent position, averaged over layers, heads, and answer tokens, gives r⁺t and r⁻t; selective credit is γt = [r⁺t − r⁻t]₊. Subtracting the wrong-answer readout discounts attention shared by both outcomes, such as formatting and transitions. A uniform floor (η = 0.3) keeps supervision alive when no valid wrong answer exists.

The weights are detached with stop-gradient. Left differentiable, the optimizer lowers the weight on hard positions instead of improving their evidence margin; removing the detachment costs 1.1 points. The evidence loss enters with weight 0.2; Stage 2 runs 100 steps on a ViRL39K + Visual-CoT mixture.

Results

Five benchmarks (MMVP, BLINK, HRBench-4K/8K, MME-RealWorld), means over three seeds on Qwen2.5-VL-7B:

ReaLVR beats LVR-SFT on all five tasks, +8.4 on MMVP and +3.3 on HR-8K. Gains hold across backbones: 61.2 on Qwen3-VL-8B, 65.2 on Qwen3-VL-30B (+1.1 over LVR-RL), 58.1 on InternVL3-8B (+3.1), 41.6 on Gemma-3-12B (+2.6). On Qwen3-VL-235B-A22B it scores 81.9/75.4/71.0 on MMVP/BLINK/MME-RealWorld, above both LVR-SFT and direct decoding; training used 800 AMD MI250X accelerators, and the paper claims this is the first latent visual reasoning trained at that scale.

The mechanistic evidence carries more information than the scores. Replacing the eight most answer-attended latent tokens with the rest frozen drops ReaLVR's correct-answer probability from 0.70 to 0.59, versus 0.67 to 0.63 for LVR, so answer dependence on latent states roughly doubles. Target-region attention reaches about twice the background level in middle layers, versus about 1.6 for Monet and at most 1.3 for LVR. Latent distance after edits rises to 0.13–0.34. A latent-length sweep puts K = 8 highest on average, with K = 0 costing 3.0 points.

One more finding: a vocabulary head reads all 376 sampled latent states as “<” with probability 1.0 (a closing-tag prior from Stage 1), yet a linear probe recovers the BLINK task label from the mean latent at 99.9%. The states hold information, not language.

Why it matters

The diagnostic toolkit (counterfactual edit sensitivity, fixed-context replacement, region enrichment) answers “is latent reasoning actually looking at the image”, and shows why final-answer accuracy hides the failure. The fix is cheap to adopt: one contrastive loss added to GRPO, no architecture or inference change, verified from 7B to 235B across three model families, so any two-stage LVR pipeline can take it directly.

The accuracy gain over the strongest baseline is modest, +0.8 over ILVR-Stage2. The stronger result is mechanistic, and it moves in step with the scores.

Limitations

• Positive prototypes depend on ROI annotations and degrade to whole-image targets when absent; how much that costs on annotation-free data is untested.
• Enrichment curves for correct and incorrect answers nearly overlap, and the paper concedes that spatial alignment does not explain correctness; the causal chain from grounding to right answers is not closed.
• The 235B evaluation covers three benchmarks with no five-task mean, so the frontier-scale claim rests on a narrow base.
• Stage 2 runs only 100 steps for all methods, so every conclusion sits in the very early regime of RL, and all five benchmarks are perception-heavy; longer reasoning chains such as math or charts are untested.
• The main text reports LVR's prediction-change rates but not ReaLVR's overall rate, which lives in an appendix table.

Terms

• LVR (latent visual reasoning): multimodal reasoning in continuous latent tokens instead of textual chain-of-thought.
• Latent evidence-credit gap: the missing link between an outcome reward and what evidence each latent token should preserve.
• GRPO: group-relative policy optimization, an RL algorithm scoring a group of sampled completions against each other.
• Teacher forcing: feeding a reference sequence token by token during training instead of letting the model generate.
• Fixed-context replacement: swap individual latent tokens while freezing the rest, then measure the answer-probability change.

Source

• arXiv 2609.34563v1
• PDF
• Hugging Face
• GitHub · https://github.com/xixiaouab/ReaLVR-code

Related papers

• Discrete codes first: DeepMind scales autoregressive image models to coherent 256×256 generation
• BrickWright: a LEGO SDK and validator take agents from 9–69% to 100% physical validity
• Best of 22 VLMs hits 58.5% on AgentVQA as visual-state errors dominate
• LoRA on 33B MiniMax-H3 cuts missed entity-count changes from 34.7% to 21.3%
• Qwen3-VL-8B training lifts RewardData artifact Finding F1 from 3.38 to 35.42
• RTAformer gains 4.8 SGDET R@10 on transition frames over matched OED

All paper explainers