Summary

arXiv:2602.00276 — submitted 30 January 2026. "Localizing and Correcting Errors for LLM-based Planners".

Proposes Localized In-Context Learning (L-ICL): identifies the first constraint-violating step in a failed plan trace and injects a minimal corrective input-output example at that precise step — raising valid plan rate from 59% to 89%.

Applicability to Zeph

HIGH — Core idea directly addresses context corruption risk in Zeph's agent loop. When a tool call fails, the LLM currently appends a raw error blob to context. L-ICL's step-localization principle suggests injecting a structured tool_result block with only the minimal corrective signal at the failure point, preventing the context from accumulating noisy error traces across retries.

Directly applicable to zeph-core context assembly and the retry logic tracked in #2199 and #2203.

Implementation Sketch

• In native.rs/legacy.rs error handling: instead of persisting the full error trace, localize to the failing tool call and inject a structured corrective example
• Complements AgentDebug (research(reliability): AgentDebug — structured corrective feedback on tool failures to prevent context corruption (arXiv:2509.25370) #2199) structured corrective feedback architecture
• Potential to reduce cascading 400 errors (similar to the bug(tools): parallel tool call with permanent error drops tool_result, causes 400 Bad Request on next LLM turn #2197 root cause)

References

• https://arxiv.org/abs/2602.00276
• Related: research(reliability): AgentDebug — structured corrective feedback on tool failures to prevent context corruption (arXiv:2509.25370) #2199 (AgentDebug), research(reliability): 12-category tool invocation error taxonomy for targeted retry/fallback strategies (arXiv:2601.16280) #2203 (tool error taxonomy)