Diagnosis Is Not Prescription: Linguistic Co-Adaptation Explains Patching Hazards in LLM Pipelines

2026-05-22 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

A study reveals the "Diagnostic Paradox" in multi-module LLM agents, demonstrating that the module most responsible for failure is not always the optimal intervention point. Causal analysis consistently identified the routing module (M3) as the primary bottleneck across three agent families: gpt-4o-mini, Llama 4 Scout, and Qwen3-32b. However, injecting prompt-level correction examples into M3 consistently degraded performance. Conversely, patching an upstream query-rewriting module (M1) reliably improved outcomes, with statistical significance on two agent families. This effect is specific to correction-injection patching, as other M3 repair strategies (instruction rewriting, model upgrade) were neutral. The "Linguistic Contract hypothesis" explains this: downstream modules implicitly adapt to upstream error distributions, so correcting the bottleneck breaks this alignment. A per-agent co-adaptation measure consistently associated higher co-adaptation with patching harm across all three agent families.

Key takeaway

For Machine Learning Engineers debugging multi-module LLM agents, directly patching the identified bottleneck module with prompt-level corrections can severely degrade performance. Instead, you should prioritize applying corrections to upstream modules, such as query rewriters, to improve overall agent outcomes. Measure linguistic co-adaptation between modules to predict potential patching hazards and avoid disrupting implicit downstream alignments.

Key insights

Correcting LLM agent bottlenecks directly can degrade performance due to downstream modules' implicit adaptation to upstream errors.

Principles

• Diagnosed LLM agent bottlenecks are not optimal patch targets.
• Downstream modules implicitly adapt to upstream error distributions.
• Correcting bottlenecks can break implicit linguistic alignment.

Method

The Causal Intervention-based Analysis (CICA) framework measures failure index F, causal contribution ΔFi, Natural Indirect Effect (NIE) for per-task fates, and applies Correction-Correction Patching (CCP) with module-normalized routing.

In practice

• Prioritize upstream modules for prompt-level corrections.
• Avoid direct prompt-level corrections on routing modules.
• Measure co-adaptation to predict patching hazards.

Topics

• LLM Agents
• Causal Intervention Analysis
• Prompt Engineering
• Linguistic Co-Adaptation
• Debugging Pipelines
• τ-bench

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• 4 Layer "AI Harness" For LLMs (+54%). Really? — The "AI Harness" improves LLM agent performance by addressing interface mismatches and syntax issues, not by modifying the frozen model.
• Not All Errors Are Equal: A Systematic Study of Error Propagation in Large Language Model Inference — The study reveals how soft errors propagate in LLM inference, identifying vulnerabilities and proposing software-only reliability improvements.
• Modular Monolingual Adaptation using Pretrained Language Models — Modular PLM adaptation, freezing embeddings while tuning the rest, improves low-resource language performance.
• Towards Knowledge Alignment in Code LLMs: Contrastive Unlearning for Evolving APIs — CURE uses contrastive unlearning to both suppress deprecated APIs and promote valid replacements in Code LLMs.
• The Hidden Reason LLMs Fail in Conversations: CCOPD — CCOPD re-anchors LLM responses in multi-turn conversations to a canonical context, mitigating "self-anchored drift" from piecemeal information.
• Weight Patching: Toward Source-Level Mechanistic Localization in LLMs — Weight Patching identifies where LLM capabilities are encoded in parameters, distinguishing source carriers from aggregation modules.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.