Paper Maps Token-Level Attention in MLLMs, Steers It Test-Time

Most interpretability work on multimodal language models has been spatial: which layer, which head, which circuit activates for visual versus linguistic information. A new preprint on arXiv titled "Attending to Multimodal Generation One Token at a Time" flips the question to temporal, looking at how attention shifts token by token as a model produces a response, and reporting that the shifts are patterned enough to steer.

The authors track four open-weight multimodal LLMs across two mainstream model families as they generate autoregressively. Three consistent patterns come out. Attention to the input image peaks on the tokens that require image-derived information. Instruction tokens are revisited during task transitions, as if the model is checking its brief when the response has to switch context. And attention to previously generated tokens climbs steadily as generation proceeds, so the model leans more on its own prior output the further it gets in.

To test whether those peaks are actually doing work rather than being decorative correlations, the paper runs causal attention-blocking interventions and profiles what breaks. The described failure modes are specific: responses fall back to language priors, exhibit cross-modal leakage, deny, or recover. That is a menu of degradations rather than a single collapse, which is what you want to see if you are going to claim a mechanism.

The pragmatic payoff is at the end. Informed by that dynamics profile, the authors propose a simple test-time intervention that boosts attention to the relevant modality at the right time, and report it significantly improves multimodal task performance. No retraining, no new weights, no data collection. If it replicates on the open-weight stacks teams actually deploy, it belongs in an inference-server config rather than a fine-tuning budget.

The honest caveat is what the abstract does not give you. The specific four models, the benchmarks, and the size of that improvement all sit in the paper body rather than the summary retrieved for this write-up. The work is also confined to open-weight MLLMs, so whether the same dynamics hold for closed frontier systems is an open question, and attention-blocking as evidence of function is a contested proxy. Read the mechanistic claims as suggestive.

For teams working with open-weight vision-language systems, the direction is the part worth watching. If attention has a schedule, decoding can be scheduled around it, and that is a lever you own without touching weights.