Research map / Multi-agent systems and games / LLM agents
Recursive Semantic Divergence for LLM Agent Consistency
Harshavardhan Abichandani, Penny Chong, Atin Ghosh, Daniel Dahlmeier
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Studies run-to-run inconsistency of tool-using LLM agents from identical starting states and proposes a recursive semantic divergence measure for it, based only on the abstract's framing.
Problem
LLM agents with tool use behave inconsistently across independent runs from the same initial state (different tool calls, contradictory outputs, divergent strategies), and existing metrics that score task completion do not capture this consistency.
Approach
Proposes a recursive semantic divergence measure to quantify agent behavioral consistency across runs, as opposed to completion-based correctness scores. Further details are not available in the provided abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune with RSD reward (GRPO) | Qwen-2.5-3B · ToolSandbox | AUC CV: 0.90 -> 0.44 | Improved |
| fine-tune with RSD reward (GRPO) | Qwen-2.5-7B · ToolSandbox | AUC CV: 0.38 -> 0.28 | Improved |
| fine-tune with RSD reward (GRPO) | Qwen-2.5-3B · tau2-bench | AUC CV: 1.32 -> 0.91 | Improved |
| fine-tune with RSD reward (GRPO) | Qwen-2.5-7B · tau2-bench | AUC CV: 0.78 -> 0.89 | Hurt |
| minimize token entropy as reward | Qwen-2.5-7B · ToolSandbox | AUC CV: 0.38 -> 0.50 | Hurt |
| minimize token entropy as reward | Qwen-2.5-3B · ToolSandbox | AUC CV: 0.90 -> 0.50 | Improved |
| minimize semantic entropy as reward | Qwen-2.5-7B · tau2-bench | AUC CV: 0.78 -> 0.78 | No clear effect |
| SFT on expert trajectories | Qwen-2.5-7B · ToolSandbox | AUC CV: 0.38 -> 0.45 | Hurt |
| SFT on expert trajectories | Qwen-2.5-3B · tau2-bench | AUC: 0.200 -> 0.289 | Improved |
| mix progress delta reward into RSD reward (lower lambda) (ablation) | RSD fine-tuned Qwen-2.5-7B · ToolSandbox | AUC CV: 0.28 -> 0.36 -> 0.37 -> 0.42 | Hurt |
| mix progress delta reward into RSD reward (lower lambda) (ablation) | RSD fine-tuned Qwen-2.5-7B · ToolSandbox | AUC: 0.555 -> 0.533 -> 0.532 -> 0.518 | Hurt |
| fine-tune with RSD reward (GRPO) | Qwen-2.5-7B · ToolSandbox (direct RSD eval) | RSD score: 0.672 -> 0.317 | Improved |
Closest papers by meaning
- Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Use (ICML 2026 · similarity 0.88)
- Test-Time Adaptation for LLM Agents via Environment Interaction (ICLR 2026 · similarity 0.88)
- Gaia2: Benchmarking LLM Agents on Dynamic and Asynchronous Environments (ICLR 2026 · similarity 0.88)
- FT-Dojo: Towards Autonomous LLM Fine-Tuning with Language Agents (ICML 2026 · similarity 0.88)
- From Words to Actions: Unveiling the Theoretical Underpinnings of LLM-Driven Autonomous Systems (ICML 2024 · similarity 0.88)
- Learning Agentic Policy from Action Guidance (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
