# Deep Research Agent

CORE, a training-free harness from Qualcomm AI Research, splits retrieval into coarse and fine stages: a dense retriever returns whole passages, then token-level similarities from a late-interaction reranker are reused to hand the agent only the most relevant snippet from each passage. In controlled closed-search comparisons over fixed corpora, CORE reaches the highest answer accuracy and evidence recall in every model-benchmark group across three agent models and three benchmarks, with corpora ranging from 100K web documents to 553M ClimbMix documents.

CORE, a training-free harness from Qualcomm AI Research, splits retrieval into coarse and fine stages: a dense retriever returns whole passages, then token-level similarities from a late-interaction reranker are reused to hand the agent only the most relevant snippet from each passage. In controlled closed-search comparisons over fixed corpora, CORE reaches the highest answer accuracy and evidence recall in every model-benchmark group across three agent models and three benchmarks, with corpora ranging from 100K web documents to 553M ClimbMix documents.

Reflective prompt evolution rewrote the two system prompts of a multi-agent deep-research agent, and under a split with no topic shared across train, validation, and test, held-out citation reward rose from 0.702 to 0.781, a gain of 0.079 that replicated across three seeds. Nearly half of the winning prompt by word count is training answers copied into the prompt, yet a controlled ablation shows this content earns no reward even on the 24 exact training questions it encodes, so the validation-minus-held-out gap fits small-sample selection under a noisy judge rather than memorization. Decomposed per-dimension feedback is what makes the gain transfer, and collapsing the reward to a single scalar drives held-out transfer negative.

Reflective prompt evolution rewrote the two system prompts of a multi-agent deep-research agent, and under a split with no topic shared across train, validation, and test, held-out citation reward rose from 0.702 to 0.781, a gain of 0.079 that replicated across three seeds. Nearly half of the winning prompt by word count is training answers copied into the prompt, yet a controlled ablation shows this content earns no reward even on the 24 exact training questions it encodes, so the validation-minus-held-out gap fits small-sample selection under a noisy judge rather than memorization. Decomposed per-dimension feedback is what makes the gain transfer, and collapsing the reward to a single scalar drives held-out transfer negative.