Looking Is Not Picking: An Attention-Segment Account of Tool-Selection Failures in LLM Agents

2026-03-26 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, long

Summary

A new analysis of LLM agent tool-selection failures reveals that models frequently attend to the correct tool but still mis-select it, refuting the common "crowded-harness" or "lost-in-the-middle" hypothesis. On real BFCL failures, the model attends most to the gold tool 80% of the time (compared to a 21% chance), with the correct tool being under-attended in only 10% of cases. This research localizes the failure to the decision readout, not the input harness. While input-side prompt repairs recover at most 23% of failures, readout-side interventions achieve 59-91% recovery. The study demonstrates that two distinct gold-pointed interventions, an additive attention-logit bias and a residual-stream steering vector, recover largely the same failures (Jaccard 0.865 pooled). Furthermore, a training-free, gold-free selector based on per-segment attention improved function-name selection by +11.9 points on BFCL and +14.9 points on Seal-Tools across five single-turn models, though it does not yet transfer to multi-turn contexts. The fix is specifically for function selection, not arguments.

Key takeaway

For AI Engineers optimizing LLM agent tool-selection, recognize that models often "see" the correct tool but fail at the decision readout. Focus your efforts on readout-side interventions, such as applying an attention-logit bias or a residual-stream steering vector, which significantly outperform prompt-based repairs. Consider implementing the training-free, gold-free attention-based selector for single-turn tool-calling tasks to boost accuracy by over 10 percentage points, but be aware it currently does not transfer to multi-turn scenarios.

Key insights

LLM tool-selection failures stem from decision readout, not attention, despite models often attending to the correct tool.

Principles

• LLM tool selection failures are a readout bottleneck.
• Attention to tools is a steerable control signal.
• Input prompt changes are less effective than readout interventions.

Method

Harness Attention Allocation (HAA) measures raw per-segment attention mass. An attention-logit bias or residual-stream steering vector can intervene at the readout. A confidence-gated selector uses per-segment attention.

In practice

• Implement readout-side interventions for tool selection.
• Use attention-logit bias for causal control.
• Apply a training-free, gold-free selector for single-turn tasks.

Topics

• LLM Agents
• Tool Selection
• Attention Mechanisms
• Readout Bottleneck
• Function Calling
• Model Interpretability

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

Related on AIssential

• Looking Is Not Picking: An Attention-Segment Account of Tool-Selection Failures in LLM Agents — LLM agent tool-selection failures stem from decision readout, not insufficient attention to the correct tool.
• To Call or Not to Call: A Framework to Assess and Optimize LLM Tool Calling — LLM tool-calling decisions, especially for web search, often misalign with true necessity and utility.
• Build an LLM from Scratch 3: Coding attention mechanisms — Self-attention mechanisms enable LLMs to selectively process input sequences, crucial for generating contextually relevant outputs.
• When Attention Closes: How LLMs Lose the Thread in Multi-Turn Interaction — LLM multi-turn instruction-following degradation stems from attention channel closure, with residual stream encoding determining post-closure behavior.
• 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.
• ToolChoiceConfusion: Causal Minimal Tool Filtering for Reliable LLM Agents — Reliable LLM agent tool selection hinges on causal necessity, not just semantic relevance or executability.

Open in AIssential →

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