When Does Critique Improve AI-Assisted Theoretical Physics? SCALAR: Structured Critic--Actor Loop for Agentic Reasoning

Authors: Vasilis Niarchos, Constantinos Papageorgakis, Alexander G. Stapleton et al.

arXiv · PDF

Summary

The authors built SCALAR, an Actor–Critic–Judge pipeline for theoretical physics problems, to study how multi-turn dialogue between a researcher-like Actor and a Critic affects solution quality. They found that multi-turn always beats single-shot attempts, but the best feedback strategy (constructive, lenient, strict, or adversarial) depends heavily on the Actor–Critic pairing—constructive feedback helps most when a lightweight Actor is guided by a stronger Critic, while same-family pairings show weaker strategy effects and sometimes favor lenient feedback.

Main takeaways:

• Multi-turn dialogue consistently improves over single-shot across all Actor–Critic pairings tested on quantum field theory and string theory problems.
• Feedback strategy matters most in asymmetric pairings (weak Actor + strong Critic), where constructive feedback improves mean scores.
• In same-family pairings (e.g., DeepSeek-R1 8B Actor + DeepSeek-R1 8B Critic), strategy effects are weaker; lenient feedback sometimes helps, strict and adversarial don't.
• Scaling within one family (8B to 70B DeepSeek-R1) improves easier problems but doesn't remove the hardest bottleneck.
• The mechanism of improvement and value of prompting choices depend strongly on the specific Actor–Critic combination, not uniformly across all setups.

Relevance

Tangentially relevant to my work on installation-path equivalence and finding prompt equivalents to fine-tuning. This explores how different interaction structures (personas, feedback strategies) affect outputs in multi-turn settings, which parallels my question of whether different installation methods (prompt vs. steering vs. fine-tuning) produce equivalent behaviors. The finding that strategy effects depend on the pairing echoes my observation that installation success varies by context.

Abstract

arXiv:2605.06772v1 Announce Type: new Abstract: As large language models (LLMs) show increasing promise on research-level physics reasoning tasks and agentic AI becomes more common, a practical question emerges: How does the interaction between researchers and agents affect the results? We study this using SCALAR (Structured Critic--Actor Loop for AI Reasoning), an Actor--Critic--Judge pipeline applied to quantum field theory and string theory problems. The Actor proposes solutions, the Critic provides iterative feedback, and an independent Judge evaluates the transcript against reference solutions. We vary the Actor persona, the Critic feedback strategy, and the Actor model family and scale. Multi-turn dialogue improves over single-shot attempts throughout, but both the mechanism of improvement and the value of different prompting choices depend strongly on the Actor--Critic pairing. Increasing the scale within one model family (e.g. from the 8B-parameter DeepSeek-R1 variant to DeepSeek-R1 70B) improves some easier-problem behavior, but does not remove the hardest bottleneck we observe. Critic feedback strategy matters most clearly in the asymmetric Actor--Critic setting (e.g., a lightweight Haiku Actor guided by a stronger Sonnet Critic), where constructive feedback improves mean-score outcomes. In same-family Actor--Critic settings, strategy effects are weaker: lenient feedback is sometimes favored, while strict and adversarial feedback are not beneficial. Taken together, SCALAR provides a controlled testbed for evaluating which interaction structures help or hinder AI-driven scientific discovery.