Skip to main content
Back to timeline
arXivSource publication:

Agents now hand off latent states, not just answers: LProp tops eight settings while using the fewest inference tokens

Synopsis

When a large language model agent passes an intermediate judgment downstream without its uncertainty, later steps can treat a tentative conclusion as fact. LProp hands off a latent representation of each candidate answer together with its probability mass, then revises the inherited belief in latent space as new evidence arrives, instead of passing a scalar score or a long text description. On BoolQ, StrategyQA, MMLU mathematics and HotpotQA with Qwen3-8B and Mistral-7B-Instruct-v0.3 backbones, it reaches the highest AUROC, AUARC and F1 in all eight settings while using the lowest mean inference tokens.

Source-provided article image: LProp: Propagating Latent Uncertainty through Evidence-Conditioned Belief Revision in LLM Agents
arXiv

Interpretation

An agent handing a judgment to the next step can now pass each candidate answer's latent state along with it, rather than only a confidence score or a text description. Passing scalar scores such as semantic entropy or verbalized confidence conveys how uncertain a result is but drops the alternative hypotheses later revision needs; passing structured text descriptions instead builds a growing history that raises token cost and inference time. Across four tasks, BoolQ, StrategyQA, MMLU mathematics and HotpotQA, and two backbones, Qwen3-8B and Mistral-7B-Instruct-v0.3, LProp has the highest AUROC in all eight settings, with gains over the strongest baseline in each setting ranging from 0.0816 to 0.2752.

When new evidence arrives, the inherited answer distribution is rewritten directly, without extra trajectory sampling or generated textual reflection. SAUP, TRACER and UProp mainly assess how much propagated uncertainty there is, while AUQ stores confidence and explanations in memory and triggers reflection before execution, where stored assessments can go stale or inconsistent and reflection adds context-processing and generation cost. In the same eight settings LProp has the highest F1 and the lowest mean inference-token usage, with unweighted mean per-setting reductions of 61.2% relative to AUQ and 71.1% relative to UProp.

The latent content of candidate answers contributes on its own, beyond answer probabilities or receiver model capacity. Matched feature ablations replace the handoff representation with probability-only or 128-token text summaries, replace latent content with frozen content or Gaussian noise, and replace the receiver with a parameter-matched MLP or log-linear fusion. With Qwen3-8B frozen, full LProp has the highest final correctness on all four benchmarks and recovers more initially wrong answers than either alternative handoff representation; on StrategyQA, frozen content recovers slightly more errors but introduces substantially more regressions.

Perspective

The demonstrated scope is a controlled source-receiver episode on four question-answering and mathematical-reasoning benchmarks: the source builds candidate support from private evidence, and the receiver gets the carrier, the question and later public evidence without the source's raw private context. The method requires white-box access to backbone hidden states and to the answer scores used to initialize candidate support, so models exposed only through text-generation or token-probability interfaces cannot directly supply the required latent features. The efficiency result is measured under the stated token-accounting protocol and is not a hardware-independent speedup.

The efficiency analysis reports inference-token usage under the specified accounting protocol rather than hardware-independent speedups; hidden-state extraction, carrier storage and learned updates add computational and storage requirements that token counts do not fully characterize, and actual latency and device memory depend on hardware, model size, batching and serving implementation. The latent interface is tied to the backbone's representation space, and source and evidence projections, feature normalization and learned heads are fitted per setting, so changing the backbone or its hidden-state representation may require refitting. On HotpotQA, recovery is restricted to correct answers already present in the source support.