Sign InOpen Brain
arXivPaperNeeds Review

The Rise of Verbal Reinforcement Learning

This survey separates verbal feedback by when it acts and what it changes, helping agent builders distinguish task specification, test-time guidance, and parameter updates.

arXiv · Sep 1, 2026
Open Source Open MarkdownOpen JSON
Source Summary

The survey defines **Verbal Reinforcement Learning** as improving agents through natural-language feedback. It organizes prior work by when feedback takes effect and what it changes, producing **three pillars** spanning task definition, inference-time reasoning, and training.

Practical Implication

Use the taxonomy to specify your feedback loop before implementing it: decide whether language should define goals and rewards, steer deliberation during a run, or become training data that changes parameters. Those choices imply different storage, evaluation, and update mechanisms.

Agent-Ready Context
The survey defines **Verbal Reinforcement Learning** as improving agents through natural-language feedback. It organizes prior work by when feedback takes effect and what it changes, producing **three pillars** spanning task definition, inference-time reasoning, and training.

Use the taxonomy to specify your feedback loop before implementing it: decide whether language should define goals and rewards, steer deliberation during a run, or become training data that changes parameters. Those choices imply different storage, evaluation, and update mechanisms.

This is a unified account of an emerging field, not a benchmark establishing which pillar works best. The abstract outlines categories, challenges, and opportunities but provides no comparative performance or deployment guidance.
Connected Context · Feed7 Judgment

This supplies a design vocabulary for feedback-driven agents: language can specify the task, alter reasoning within a run, or update the model afterward. Against prior harness and memory work, it clarifies that retaining feedback as context is only one intervention class and should not be conflated with parameter learning. It organizes implementation choices but does not resolve which feedback mechanism best addresses a given reliability gap.

A Taxonomy of Cognitive Capability Gaps in Generative and Agentic AIThe capability-gap taxonomy identifies what an agent lacks, while the three-pillar framework identifies where language feedback could intervene; together they separate diagnosis from the choice of update mechanism.The Unreasonable Effectiveness of Separating the Task from the Model — Maxime Rivest & Isaac MillerDSPy’s separation of task contracts from prompts and harnesses aligns with distinguishing language that defines goals from language that steers inference or becomes training data.The Era of Compound Engineering — Kieran Klaassen, Every/CoraCompound engineering makes reviewed corrections reusable as future context, providing a concrete inference-time or memory-based feedback loop that the survey distinguishes from parameter updates.
Context Map
agentresearch#prompting#agent-memory#harness-engineering
Uncertainty
This is a unified account of an emerging field, not a benchmark establishing which pillar works best. The abstract outlines categories, challenges, and opportunities but provides no comparative performance or deployment guidance.