Structured but Silent: Probing Capability Requirements in LLM Hidden States
In this paper, we investigate whether these query-side capability requirements are linearly decodable from LLM hidden representations prior to generation, and how this hidden-state accessibility compares with explicit verbal classification.
Key points
- Reliable tool use requires more than triggering a mechanism or matching a query to an API description.
- We introduce TACIT, a framework that decomposes external requirements along three fundamental axes: Source, Transformation, and World Effect, defining eight structurally distinct capability classes.
- Using 1,600 balanced training queries from benchmarks, synthetic examples, and new domain scenarios, we train linear probes on pre-generation hidden states from four open-weight LLM families.
- Our empirical results demonstrate that fine-grained capability structures are linearly decodable with high accuracy across all models.
Sources (1)
- [1]Structured but Silent: Probing Capability Requirements in LLM Hidden StatesarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 6, 09:12 AM
In this paper, we investigate whether these query-side capability requirements are linearly decodable from LLM hidden representations prior to generation, and how this hidden-state accessibility compares with explicit verbal classification.
Reliable tool use requires more than triggering a mechanism or matching a query to an API description.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 6, 2026CADFather Reconstructs Parametric CAD Programs Using Coordinated Tools
- Sep 28, 2026Holo4: powering generalist computer-use agents
- Sep 24, 2026Introducing Gemini 3.8 Live with Live Avatar
- Sep 15, 2026Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking
- Aug 26, 2026Intelligent transcription with Gemini 3.5 Transcribe
- Aug 13, 2026Introducing Gemini 3.7 Flash