RL4LLM-Agents

We invite submissions on Reinforcement Learning for LLM-based agents along either axis, or across both. We especially welcome work grounded in finance, trading, risk management, and related domains. Topics of interest include, but are not limited to:

Track A: Model-side RL (weight space)

• RL-based fine-tuning of LLMs: RLHF, RLAIF, PPO, GRPO, DPO, and other RL-from-feedback methods for shaping agent behavior.
• RL with verifiable rewards and process reward models for multi-step reasoning and tool use.
• Reward modeling and shaping, including sparse-reward settings and reward-hacking mitigation.
• Exploration, credit assignment, and sample efficiency when training agents over long horizons.
• Theoretical foundations: convergence, generalization, and optimization in RL for language models.

Track B: Harness-side RL, or in-context RL (context space)

• In-context reinforcement learning and test-time learning that improves an agent without any weight update.
• Reflection as a learning signal: self-critique, verbal feedback, and trajectory-level error analysis in the role of the gradient.
• Memory and skill libraries: generation, extraction, retrieval, consolidation, versioning, and forgetting.
• Automatic prompt, instruction, and program optimization, including compiled-pipeline and search-based approaches.
• Self-generated curricula and open-ended skill discovery in interactive environments.
• Tool and skill synthesis: agents that build, name, and hierarchically organize their own capabilities.
• Validation gates and lifecycle control: preventing context pollution, quantifying degradation from unvetted memory, and deciding what is allowed to persist.
• Evolutionary and population methods over harness configurations, prompts, and skill sets.
• Transfer of harness artifacts: strong-to-weak skill teaching, and distilling learned context into smaller or cheaper models.

Cross-cutting topics

• Weights or context: when to move which, and the cost, latency, data, and governance trade-offs involved.
• Hybrid pipelines: harness search that generates training data for weight-space RL, and fine-tuned models that shrink the harness they need.
• LLM agents for sequential decision-making: planning, reasoning, and acting in multi-step environments.
• Multi-agent LLM systems: cooperation, competition, and emergent behavior among multiple LLM-based agents.
• Agentic RAG and tool use: agents that retrieve information, call APIs, and interact with external environments.
• Evaluation benchmarks, datasets, reproducibility, and observability for agents that learn along either axis.
• Safety, alignment, auditability, and robustness of learning agents in high-stakes domains.
• LLM agents in financial applications: algorithmic trading, portfolio management, market simulation, risk assessment, and research automation.

Download the Call for Papers (PDF)

We welcome the following types of submissions:

• Full papers (up to 8 pages, plus up to 2 additional pages for references): presenting mature research with rigorous evaluation.
• Short papers and extended abstracts (at least 4 pages, plus up to 2 additional pages for references): presenting early-stage work, position papers, or proofs-of-concept.
• System and reproducibility reports (at least 4 pages, plus up to 2 additional pages for references): open implementations, benchmarks, or negative results along either axis.

Submissions should be formatted using the ACM proceedings template and submitted as PDF. All submissions will be peer-reviewed. Accepted papers will be presented at the workshop (oral or poster, based on reviewer recommendations).

Length guidance from ICAIF 2026: workshop papers should be at least four pages, plus up to two additional pages for references.

Submission portal: submissions are handled through Microsoft CMT. Please select the track “Workshop on Reinforcement Learning for LLM-based Agents (RL4LLM-Agents)”.

Submit via CMT

Proceedings: this workshop is non-archival. Accepted papers will not appear in the ACM ICAIF 2026 proceedings, so authors remain free to submit extended versions elsewhere. Accepted papers may be posted on this website.

Submissions must describe original, unpublished work and must not be under review elsewhere. We follow a single-blind review process.

Krati Saxena

Fabrizio Dimino

Abhinav Arun

Jacopo Biggiogera

Additional program committee members to be announced.