Researchers at the University of Southern California trained a separate Qwen3-8B model to advise Gemini 3.7 Flash and Claude Sonnet 4.6 on tool-use tasks. Before each response, the advisor recommends a next action or abstains; post-task feedback trains the advisor, not the two task-performing models.
Gemini and Claude are the executors: they produce each answer and tool call, while Qwen3-8B supplies turn-by-turn guidance. The paper calls them frozen because their model weights do not change.
After an interaction, a reflector uses the executor’s replies, tool results and task checks to suggest corrections. To decide which advice turns merit further training, AdviSD scores the same recorded executor reply twice with the advisor: one prompt includes the original advice and the other removes it. A large difference in the reply’s likelihood score flags that decision for extra training. A feedback-informed copy of the advisor then teaches the training model to produce the selected advice from the original context alone. This runs during training and requires neither executor token probabilities nor additional executor runs.
On BFCL-v3, the advisor scored 67.8 with Gemini 3.7 Flash and 63.1 with Claude Sonnet 4.6, compared with 61.4 and 58.9 under reward-only GRPO training. Those are gains of 6.4 and 4.2 percentage points. On EnvScaler’s graded score, AdviSD scored 84.8 versus 79.7 with Gemini and 84.6 versus 80.7 with Claude, gains of 5.1 and 3.9 score points. Each result is a mean from three training runs evaluated on held-out tasks: 320 BFCL-v3 tasks and 200 EnvScaler tasks.
Without further training, the advisors also beat the unassisted executors on an equal-weight average of four additional tool benchmarks, by 2.7 points for Gemini and 3.6 for Claude. The pattern was not universal: the Claude advisor scored 1.7 points below standalone Claude on ACEBench’s multi-step test. The study is a preprint with two executor families and one shared reflection model; updates to the API models may also make exact replication difficult. It is benchmark evidence, not a deployment result.
Sources
- AdviSD preprint and version record, submitted September 29, 2026
- Full paper: method, held-out benchmark results, transfer tests and limitations
- Lead image: Figure 1 by Rishabh Agrawal et al., reproduced under CC BY 4.0.