How Do Agentic LLMs Decide to Call Tools? A Tool-Call Vector Shaped by Suppression
cs.AI, cs.LG
Submitted: 2026-10-07
Updated: 2026-10-07
Code: https://github.com/XijieGo/MI4ToolCalling
Sources
- Toolformer: Language Models Can Teach Themselves to Use Tools
- Qwen3 Technical Report
- Gorilla: Large Language Model Connected with Massive APIs
- ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs
- Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small
- AGENTIF: Benchmarking Instruction Following of Large Language Models in Agentic Scenarios
- Program Synthesis with Large Language Models
- Evaluating Large Language Models Trained on Code
- Refusal in Language Models Is Mediated by a Single Direction
- How to use and interpret activation patching
- $\tau^2$-Bench: Evaluating Conversational Agents in a Dual-Control Environment
- Transcoders Find Interpretable LLM Feature Circuits
- Representation Engineering: A Top-Down Approach to AI Transparency
- Have Faith in Faithfulness: Going Beyond Circuit Overlap When Finding Model Mechanisms
- API-Bank: A Comprehensive Benchmark for Tool-Augmented LLMs
- $\tau$-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains
- Understanding intermediate layers using linear classifier probes
- Eliciting Latent Predictions from Transformers with the Tuned Lens
- Understanding and Mitigating Cross-lingual Privacy Leakage via Language-specific and Universal Privacy Neurons
- Attribution Patching Outperforms Automated Circuit Discovery
Related papers
- MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
- Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
- Agents' Last Exam
- The Clinician's Veto: Navigating Trust, Liability, and Uncertainty in Autonomous AI Prescribing
- MindHelper: Closed-Loop Embodied Mental-State Reasoning for Precision Intervention
- Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems