A small side model can make a top financial research agent smarter and cheaper, without touching the agent's own weights.
The approach, called hindsight structured advice distillation, trains an 8B advisor model to read each incoming query and write custom guidance that gets tacked onto the main agent's prompt. The advisor learns by hindsight: researchers run a frozen agent through several strategies on a set of graded practice queries, then have a teacher model study the grades and the agent's traces to write the advice that would have helped. The advisor then learns to produce that advice from the query alone, without ever seeing the grades or traces at inference time. Tested on FrontierFinance, a benchmark of expert financial research questions, the advisor lifted Samaya's agent, already the leaderboard leader, from a 52.9% qualification rate to 56.6%, beating the agent's own high-effort mode at 37% lower cost per query.
That matters because the standard fix for a capable-but-detail-blind agent, like one that doesn't know which SEC filing holds which figure, is either retraining it or cramming every edge case into a static prompt that's identical for every query. Retraining is expensive and often impossible on closed or shared systems, and static prompts still can't tailor themselves to each input. A cheap, separately trained advisor sidesteps both problems, and the same technique reportedly helped on WebShop, BIRD text-to-SQL and ExpertLongBench too.
Still, a 3.7-point bump on one benchmark, beating a slower high-effort setting that was presumably never built for cost efficiency, is a modest claim dressed in a mouthful of an acronym, and worth watching if it holds up outside expert-curated test sets.