Advances in multimodal large language models have boosted multimodal embeddings. Recent work augments inputs with self‑generated reasoning chains, yielding richer embeddings. However, conventional retrieval supervision only provides pairwise relevance and does not explicitly guide the reasoning used for augmentation. Reasoning is unobserved and non‑unique; a single input can admit many plausible traces that differ in retrieval utility. Existing methods optimize reasoning conditioned solely on the input, limiting exploration of retrieval‑relevant traces.
We introduce the PILOT framework, treating reasoning as a latent variable and instantiating its prior and posterior as two reasoning modes of a shared MLLM. The prior conditions only on the input, while the posterior leverages additional relational evidence during training to explore a broader set of traces. To transfer the posterior’s exploratory gains to inference, we derive an importance‑weighted variational objective and realize it as a prior‑posterior co‑evolution mechanism. This mechanism jointly improves both modes under a shared inference‑compatible reward and selectively transfers posterior trajectories to the prior based on their retrieval utility and prior reachability.
On the 78‑task MMEB‑V2 benchmark, PILOT outperforms the primary baseline across all 12 task groups at both 2B and 7B parameter scales, raising overall scores by 3.2 and 3.3 points respectively, with the largest gains in visual‑document retrieval (+6.1 and +6.4 points). The results demonstrate that relational supervision is most valuable as a signal shaping reasoning exploration during training, rather than merely as a post‑hoc reward.
Review