mlsanityData to 14 Oct 2026

More Is Not Always Better: Cross-Component Interference in LLM Agent Scaffolding

arXiv 2605.05716 ↗First posted 7 May 2026cs.AI cs.CLcited by 1

On the map: LLM agents and tool use → General LLM agent frameworks and benchmarks

Abstract (arXiv)

LLM agent systems are built by stacking scaffolding components (planning, tools, memory, self-reflection, retrieval) assuming more is better. We study cross-component interference (CCI): degradation when components interact destructively. We run a full factorial experiment over all 2^5=32 subsets of five components on HotpotQA and GSM8K with Llama-3.1-8B/70B (96 conditions, up to 10 seeds). The All-In system is consistently suboptimal: on HotpotQA, a single-tool agent surpasses All-In by 32% (F1 0.233 vs 0.177, p=0.023); on GSM8K, a 3-component subset beats All-In by 79% (0.43 vs 0.24, p=0.010). Optimal component count is task-dependent (k*=1-4) and scale-sensitive: at 70B, combinations that hurt at 8B provide gains, though All-In still trails the best subset. We fit a main-effects regression (R^2=0.916, adj-R^2=0.899, LOOCV=0.872), compute exact Shapley values, and find 183/325 submodularity violations (56.3%), showing greedy selection is unreliable. A three-body synergy among Tool Use, Self-Reflection, and Retrieval (INT_3=+0.175, 95% CI [+0.003,+0.351]) is reported as exploratory. CCI replicates across model families (Qwen2.5) and is robust to prompt paraphrasing. Our findings suggest maximally-equipped agent defaults should be replaced by task-specific subset selection via interaction-aware analysis.

Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Placed in General LLM agent frameworks and benchmarks, one of 16 directions in LLM agents and tool use. Builds on 22 of its 30 references that are in the map. First posted on arXiv 7 May 2026.

Builds on

22 of 30 references.

References · 22

  1. 2021Training Verifiers to Solve Math Word Problemscited by 10,109
  2. 2022ReAct: Synergizing Reasoning and Acting in Language Modelscited by 8,695
  3. 2023Reflexion: Language Agents with Verbal Reinforcement Learningcited by 4,276
  4. 2018HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answeringcited by 3,754
  5. 2023Lost in the Middle: How Language Models Use Long Contextscited by 3,455
  6. 2020Gradient Surgery for Multi-Task Learningcited by 1,256
  7. 2023DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelinescited by 832
  8. 2023State of What Art? A Call for Multi-Prompt LLM Evaluationcited by 310
  9. 2021Efficiently Identifying Task Groupings for Multi-Task Learningcited by 222
  10. 2023Inverse Scaling: When Bigger Isn't Bettercited by 207
  11. 2024Benchmarking Complex Instruction-Following with Multiple Constraints Compositioncited by 152
  12. 2025When Thinking Fails: The Pitfalls of Reasoning for Instruction-Following in LLMscited by 53
10 more
  1. 2023SHAP-IQ: Unified Approximation of any-order Shapley Interactionscited by 51
  2. 2025How Many Instructions Can LLMs Follow at Once?cited by 20
  3. 2023HarsanyiNet: Computing Accurate Shapley Values in a Single Forward Propagationcited by 17
  4. 2025ThriftLLM: On Cost-Effective Selection of Large Language Models for Classification Queriescited by 14
  5. 2024The Unreasonable Effectiveness of Eccentric Automatic Promptscited by 12
  6. 2024Explaining Large Language Models Decisions Using Shapley Valuescited by 12
  7. 2024Position Engineering: Boosting Large Language Models through Positional Information Manipulationcited by 9
  8. 2024DemoShapley: Valuation of Demonstrations for In-Context Learningcited by 3
  9. 2026On the Paradoxical Interference between Instruction-Following and Task Solvingcited by 2
  10. 2026A Regression Framework for Understanding Prompt Component Impact on LLM Performancecited by 1

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026How Should a Prompt Optimizer Spend a Tight Budget? BudgetAPO with Noise-Adaptive Evaluationcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Arbiter: Detecting Interference in LLM Agent System Prompts
  2. 2026Attention Tax, Handoff Tax: A Stylised Model of When Multi-Agent LLM Systems Help
  3. 2026Multi-agent Scaling Across Disjunctive and Compensatory Tasks
  4. 2025How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior
  5. 2025Diagnose, Localize, Align: A Full-Stack Framework for Reliable LLM Multi-Agent Systems under Instruction Conflicts
  6. 2026What Can Component-Replacement Evidence Establish? A Critical Scoping Review of Local Decisions in LLM Agents
  7. 2026Break It Down, Pass It On: Cross-Task Skill Transfer in LLM Agents
  8. 2026Architectural Design, Not Only Model Intelligence, Governs Multi-Agent LLM Performance