AI Science Tracker
EN

Actor–critic training showed a local reversal in model-size scaling

Google DeepMind · 24 Oct 2026 · computer science

Discovery

Google DeepMind

Actor–critic training showed a local reversal in model-size scaling

Source check: Source checked. Not independent scientific confirmation.

Finding: In Craftax, a larger actor–critic model outperformed a smaller one at 10¹⁴ training FLOPs, fell behind at 3×10¹⁴ FLOPs, and regained its lead by 10¹⁵ FLOPs. The result shows a local departure from monotonic scaling for the tested policy models. The authors proposed a learning-dynamics explanation based on in-game achievements.

Technology: Actor–critic reinforcement learning; empirical compute scaling

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

AI contribution
Trained actor–critic policies at different sizes and budgets generated behavior traces used to test local compute scaling.
Test and result
The experiment trained model sizes across compute levels and examined the crossover alongside achievement frequencies.
Limitation / still unknown
The crossover concerns the evaluated sizes, training setup and game environment. The proposed explanation is observational.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2610.33420v1 · https://arxiv.org/html/2610.33420v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 21 Nov 2026. Review scope: v1 actor–critic setup, §4.4 and Figure 5.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Open in the full catalog for filters, charts and related discoveries. Browse computer science discoveries.

Related catalog entries

Links use documented relationships where available, plus shared catalog context. A shared topic or model family does not establish a scientific dependency.