ResidencyRL: Training interaction policies in full multi-turn simulation

简体中文

Too costly, not reproduced (review DONE). Research-method notes only, not medical advice or local verification of clinical capability. residency-rl is a reference key; adapter: null. Method and cost reviewed on v1; v2 posted 2026-10-09 was observed but not fully compared.

Paper information

Field	Details
Paper	Pinned v1; Google's official entry
Institution	First author Valentin Liévin: Google DeepMind; collaborators include Google Research
First public release	2026-08-07 17:04:41 UTC (arXiv v1; historical backfill, not an October release)
Author code	Explicitly not released; Code Availability cites internal training infrastructure and safety restrictions. Third-party benchmarks are not the implementation.
Local adapter or method key	residency-rl
Local reproduction code	None; no runnable reproduction, clinical tool or Evolve operator.

Paper method and evidence

Static knowledge QA does not test active questioning, tool use or long-horizon decisions. The paper combines scenario generation, information-asymmetric simulations, isolated dimension-specific judges and GRPO. Hidden references belong to environment and judges, not evaluated policies. Weighted normalized dimension scores minus truncation penalty are only reward aggregation, not the complete algorithm. Section 3 · Table 2 and Eq. 1.

flowchart LR
  A[Generate hidden and public scenario] --> B[Roll out multi-turn tool interactions]
  B[Roll out multi-turn tool interactions] --> C[Isolated judges]
  C[Isolated judges] --> D[Aggregate reward]
  D[Aggregate reward] --> E[GRPO update]

Original paper figure

Figure 1 shows training and the authors' evaluation, not local experiments. From the original paper; copyright remains with the authors.

Core equation

\[ R=3\sum_{d}w_d\bar r_d-\min(3,P),\qquad w=(2,3,1,1,1,1)/9. \]

Reported results

Reported simulated, external-task and blinded results are not clinical production benefits or industrial online A/B. Prospective real-workflow validation remains necessary. No local recomputation or import of author figures as experimental records.

Local reproduction

The internal Gemini training stack uses separate simulation/judge calls and 49,870 regular, 5,000 history-taking and 2,583 adversarial scenarios, with up to 60 turns and eight tool calls per trajectory. Section 3.4. One A100/A30 or a Mac cannot faithfully replace it, and public APIs do not expose the same weight update. No paid long probe, substitute checkpoint, scripted reproduction claim, runnable code, metrics, GPU receipt, Evolve operator or retry queue.