Researchers from multiple institutions published a study examining how AI agents can collaborate to solve complex theoretical physics problems through structured critique loops.

The team developed SCALAR (Structured Critic--Actor Loop for AI Reasoning), a three-component system where an Actor AI proposes solutions, a Critic provides iterative feedback, and an independent Judge evaluates results against reference solutions. The framework was tested on quantum field theory and string theory problems.

Multi-turn dialogue consistently outperformed single-shot attempts across all tested configurations. However, the mechanism of improvement and optimal prompting strategies varied significantly based on the specific Actor-Critic pairing used.

The researchers tested different model scales within the same family, comparing 8-parameter and 70-parameter variants of DeepSeek-R1. Larger models improved performance on easier problems but failed to overcome the most challenging bottlenecks observed in the study.

Feedback Strategy Impact Varies by Setup

Critic feedback strategy proved most important in asymmetric Actor-Critic configurations, where a lightweight model like Claude Haiku acted as the Actor while a more powerful model like Claude Sonnet served as the Critic. In these cases, constructive feedback significantly improved mean scores.

Same-family Actor-Critic pairings showed weaker strategy effects. Lenient feedback sometimes performed better, while strict and adversarial feedback approaches provided no measurable benefits.

The study authors noted that increasing model scale within one family improved some behaviors but did not eliminate the hardest reasoning bottlenecks they identified.

SCALAR provides a controlled experimental framework for evaluating which interaction structures help or hinder AI-driven scientific discovery. The research addresses practical questions about human-AI collaboration as Anthropic and other labs develop increasingly capable reasoning models.

The paper was submitted to arXiv on May 7, 2026, with authors from institutions including CERN and Queen Mary University of London.