PLUME: Latent Reasoning Based Universal Multimodal Embedding

Dimensions overall score 7.0

• PLUME outperforms strong explicit-CoT UME baselines on the 78-task MMEB-v2 benchmark.result 90%
  Evidencepartial

  On the 78-task MMEB-v2 benchmark, PLUME outperforms strong explicit-CoT UME baselines

  Implicationpartial

  Explicitly stated in the abstract with a specific benchmark reference.

  Verificationpartial

  partial

• PLUME reduces reasoning from hundreds of generated tokens to fewer than 10 latent steps, delivering over 30x faster inference.result 90%
  Evidencepartial

  reducing reasoning from hundreds of generated tokens to fewer than 10 latent steps, delivering over 30x faster inference.

  Implicationpartial

  Explicitly stated in the abstract with clear numeric metrics.

  Verificationpartial

  partial

• PLUME replaces explicit chain-of-thought reasoning with a latent reasoning framework using a short autoregressive rollout of continuous latent states.method 90%
  Evidencepartial

  We propose PLUME, a latent reasoning framework that advances UME by replacing verbalized CoT with a short autoregressive rollout of continuous latent states.

  Implicationpartial

  Directly stated in the abstract as the core method.

  Verificationpartial

  partial

• PLUME introduces a semantic-anchor-guided transition adapter to steer latent rollout along different reasoning trajectories under a fixed computation budget.technical 85%
  Evidencepartial

  PLUME further introduces a semantic-anchor-guided transition adapter that steers latent rollout along different reasoning trajectories under the same fixed computation budget.

  Implicationpartial

  Directly stated in the abstract as a key technical component.

  Verificationpartial

  partial

• PLUME adopts a progressive explicit-to-latent curriculum that uses verbalized reasoning only as a temporary training scaffold, eliminating explicit CoT at inference.method 85%
  Evidencepartial

  PLUME adopts a progressive explicit-to-latent curriculum that uses verbalized reasoning only as a temporary training scaffold and gradually transfers this behavior into hidden-state computation, eliminating explicit CoT at inference.

  Implicationpartial

  Directly stated in the abstract as a key training method.

  Verificationpartial

  partial

• PLUME is especially well-suited to retrieval settings where relevant evidence is dense, structurally complex, and difficult to organize through verbalized intermediate rationales, such as video and visual document retrieval.technical 80%
  Evidencepartial

  PLUME is especially well suited to retrieval settings where relevant evidence is dense, structurally complex, and difficult to organize through verbalized intermediate rationales, such as video and visual document retrieval.

  Implicationpartial

  Directly stated in the abstract as a specific application strength.

  Verificationpartial

  partial

• PLUME may have limitations if extended to tasks that inherently require explicit intermediate reasoning steps, or if the latent steps are insufficient for complex queries.limitation 85%
  Evidencepartial

  PLUME may have limitations if extended to tasks that inherently require explicit intermediate reasoning steps, or if the latent steps are insufficient for complex queries.

  Implicationpartial

  Explicitly stated in the analysis excerpt under 'caveats'.

  Verificationpartial

  partial

• Explicit CoT incurs substantial inference overhead and can compress rich multimodal evidence into a narrow textual bottleneck.technical 80%
  Evidencepartial

  explicit CoT incurs substantial inference overhead and can compress rich multimodal evidence into a narrow textual bottleneck.

  Implicationpartial

  Directly stated in the abstract as a motivation for the work.

  Verificationpartial

  partial

Competitive landscape

PLUME offers a faster and more efficient universal multimodal embedding by using latent reasoning instead of explicit chain-of-thought, significantly reducing inference time for complex retrieval tasks.