Tokengeist: Multi-Turn Attribution Tracing in Agentic Conversations
Researchers introduced multi-turn context attribution (MTCA) to trace how tokens from prior turns shape a language model's response in multi-turn conversations. Existing methods process context in a single pass, missing layered dependencies. Tokengeist, an attribution-method-agnostic framework, recovers full dependency paths by casting attribution as recursive traversal of a directed acyclic graph (DAG) over turns. The team will release MTCABench, a benchmark of 3,845 target spans. The paper is available on arXiv.
Key facts
- Multi-turn context attribution (MTCA) is introduced.
- Tokengeist is an attribution-method-agnostic and scalable framework.
- Tokengeist recovers full dependency paths by casting attribution as a recursive traversal of a directed acyclic graph (DAG) over conversation turns.
- Existing context attribution methods process the full context in a single pass.
- MTCABench is a benchmark of 3,845 target spans to be released.
- The paper is on arXiv with ID 2607.22610.
- The work addresses layered, non-linear structure of real-world dialogues.
- Tokengeist is designed for multi-step reasoning tasks.
Entities
Institutions
- arXiv