From Routing Signals to Selective Review: Visual regrounding in MoE VLMs

What it claims

A linear probe over router probabilities at the question's target-object tokens predicts, before any answer is generated, whether the queried object is missing from the image. The probe's output decides whether to re-prompt the unchanged model with a "verify the object is present" instruction. Reported end-to-end gains are in percentage points (pp):

The Gemma gain on the external OBER dataset is close to nothing. The paper also says the signal sits on the target-object token, appears in early MoE layers, and is spread across experts that can partly substitute for one another.

How it works

1. A frozen rule-based extractor pulls the target noun phrase from the question. It abstains on unsupported or ambiguous questions, and no gold object annotation is used at inference time.
2. The model's tokenizer maps that phrase to sub-token positions in the full multimodal sequence.
3. During normal prompt prefill, the method records the router's full pre-top-k softmax distribution, over all experts and not just the executed top-k. It averages this over the target span and concatenates it across layers. For Qwen (48 layers × 128 experts, k=8) that gives a 6,144-dim vector.
4. Each dimension is standardized using training-split statistics only.
5. An L2-regularized logistic probe, one per model, scores absence: Regularization strength is chosen by validation ROC-AUC. The decision threshold θ is chosen by validation balanced accuracy. * All of these are frozen before test.
6. If the score is at or above θ, the same image and question are regenerated with a review instruction. Otherwise the standard answer stands. Review costs one extra forward pass, paid only on triggered inputs.

flowchart LR
  A[Standard prefill] --> B[Target-token routing]
  B --> C{Probe score ≥ θ?}
  C -- no --> D[Standard decode]
  C -- yes --> E[Regenerate with target-aware review prompt]

Two review prompts were tested. "Generic" asks the model to inspect the image carefully. "Target-aware" asks it to verify that the object is present and to say so explicitly if it is absent.

Evidence

Probe training data: 20,000 train, 5,446 validation and 5,776 test pairs. Each pair is one target-present image and one target-absent image. Images from the same source scene stay in the same split.

Shuffled-label control (Gemma, 20 runs): ROC-AUC 0.515 ± 0.062, close to chance.

Review experiment setup:

• Manually audited sets: 120 GQA pairs and 115 OBER pairs (5 filtered out as ambiguous).
• Five sampled answers per pair per prompt.
• All answers graded by hand against the images.
• Policy accuracy counts both corrections and errors introduced by review.

How to use it

• You need access to router logits during prefill, which rules out closed or API-only MoE models.
• You need a per-model labeled present/absent training set; probes are trained separately for each model.
• You need the noun-phrase extractor.

The fetched text gives no code, repository or licence.

Limits

• Thresholds do not transfer. The ranking signal carries across datasets, but the GQA-calibrated threshold drops to 65–71% accuracy on OBER. The authors say recalibration is required for cross-dataset use.
• Weak external gain for Gemma. The +1.39 pp on OBER suggests the benefit depends on the model and the data.
• Small review sets. End-to-end gains rest on 120 and 115 manually graded pairs.
• Narrow scope. Only single-image, single-target questions where a noun phrase can be extracted. The extractor abstains otherwise.
• Not a general hallucination detector. Other visual hallucination types are explicitly excluded.
• False-positive review harm. It is described as "limited overall," with risk controlled by jointly choosing the threshold and the review prompt. The fetched text is truncated before the detailed harm, layer and expert analyses, so those figures are not reproduced here.
• Probe weights are not causal. The authors caution that probe coefficients are conditional predictive contributions, not causal expert importance.