Bayesian control for coding agents

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Theodore Papamarkou, Vladislav Smirnov, Viktor Mazanov, Artem Vazhentsev, Preslav Nakov, Timothy Baldwin, Artem Shelmanov
• Venue: arXiv: cs.AI
• Published: 2026-06-23
• Collected: 2026-06-24
• URL: http://arxiv.org/abs/2606.24453v1
• PDF: https://arxiv.org/pdf/2606.24453v1
• DOI: UNKNOWN

Plain-Language Brief

The paper proposes a Bayesian controller for coding agents that make decisions based on a belief over candidate correctness, rather than fixed rules. This method is found to be most useful when verification is costly and the critics are informative but imperfect.

What the Document Says

• Modern coding agents combine LLM generators with various tools, including cheap diagnostics and expensive verifiers.
• The decision-making in tool usage by these agents is usually governed by orchestrators using fixed rules.
• The authors propose a Bayesian control method for orchestration, where a controller maintains a belief over candidate correctness and dynamically decides whether to gather more evidence, refine the candidate, verify it, or stop.
• The study was conducted across six generators and nine coding benchmarks.
• The Bayesian control is found to be most valuable when verification is costly and critics are informative but imperfect.
• Beyond control, the belief state yields an interpretable correctness score that outperforms token-probability and raw tool-success baselines for uncertainty quantification.

Key Concepts Mentioned

• coding agents
• LLM generators
• orchestrators
• Bayesian controller
• belief over candidate correctness
• cost-sensitive sequential hypothesis testing

Claims Made by the Document

• The proposed Bayesian control method is most valuable when verification is costly and critics are informative but imperfect.
• The belief state yields an interpretable correctness score that outperforms token-probability and raw tool-success baselines.

Evidence Used

• Across six generators and nine coding benchmarks, Bayesian control proves to be most valuable when verification is costly and critics are informative but imperfect.
• Beyond control, the belief state yields an interpretable correctness score that outperforms token-probability and raw tool-success baselines for uncertainty quantification.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Modern coding agents pair LLM generators with various tools, including cheap diagnostics and expensive verifiers. The tool-use decisions are typically governed by orchestrators that often use fixed rules and ignore uncertainty. We formulate orchestration as cost-sensitive sequential hypothesis testing: a Bayesian controller maintains a belief over candidate correctness and dynamically decides whether to gather more evidence, refine the candidate, verify it, or stop. Across six generators and nine coding benchmarks, Bayesian control proves to be most valuable when verification is costly and critics are informative but imperfect. Beyond control, the belief state yields an interpretable correctness score that outperforms token-probability and raw tool-success baselines for uncertainty quantification.