Learnijoy
ResearchAI ResearchAI Engineering & DevTools

Tokengeist Traces LLM Attribution Across Multi-Turn Conversations

Jessica Tang, Shraddha Barke, Sharad Agarwal· July 28, 2026 View original

Summary

Tokengeist is a new framework that traces how language models attribute information across multiple turns in a conversation, revealing deep dependencies that single-pass methods miss. It casts attribution as a recursive graph traversal to identify which prior turns and tokens influenced a specific response.

Current methods for understanding why a language model generates a particular response often fall short in multi-turn dialogues. They typically process the entire context at once, which can obscure the complex, layered dependencies that build up over a conversation. This new research introduces the concept of multi-turn context attribution, aiming to not only identify directly relevant prior turns but also to understand how those turns themselves were influenced by even earlier context. The proposed framework, Tokengeist, tackles this by modeling attribution as a recursive journey through a directed acyclic graph representing conversation turns. This approach allows it to uncover full dependency paths, addressing what researchers call "provenance collapse" in existing single-pass methods. A new benchmark, MTCABench, with thousands of annotated conversation spans, demonstrates Tokengeist's superior performance, achieving 90% source recall compared to under 20% for flat attribution methods.

Why it matters

Understanding the provenance of LLM outputs in complex dialogues is crucial for debugging, ensuring factual accuracy, and building trustworthy AI agents, especially in critical applications.

How to implement this in your domain

  1. 1Integrate multi-turn attribution tools into AI agent development pipelines to enhance transparency.
  2. 2Utilize recursive attribution methods to debug unexpected or incorrect agent responses in conversational AI.
  3. 3Develop internal benchmarks similar to MTCABench to evaluate the explainability of proprietary conversational models.
  4. 4Train AI teams on the importance of multi-turn context understanding for robust agent design.

Who benefits

Customer ServiceHealthcareLegalFinancial ServicesAI Development

Key takeaways

  • Traditional LLM attribution methods struggle with multi-turn conversations, missing deep dependencies.
  • Tokengeist offers a novel, recursive framework for tracing attribution across conversation turns.
  • The method significantly improves source recall for multi-hop dependencies compared to single-pass approaches.
  • Understanding multi-turn attribution is vital for building more reliable and explainable AI agents.

Original post by Jessica Tang, Shraddha Barke, Sharad Agarwal

"arXiv:2607.22610v1 Announce Type: new Abstract: When a language model produces a response in a multi-turn conversation, which tokens from prior turns shaped that answer, and how did those dependencies propagate across prior turns? Existing context attribution methods process the…"

View on X

Originally posted by Jessica Tang, Shraddha Barke, Sharad Agarwal on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI ResearchAI Engineering & DevTools

StageGuard Improves Sleep Staging by Enforcing Physiological Constraints

StageGuard is a new framework that enhances automated sleep staging by integrating physiology-informed priors, ensuring that deep learning models produce hypnograms that adhere to known biological rules. It significantly reduces physiologically implausible transitions and fragmentation while maintaining or improving accuracy.

Juntang Wang, Yihan Wang, Hao Wu, Jiayu Gao, Shixin Xu, Dongmian ZouJul 28, 2026
AI ResearchAI Engineering & DevToolsAI News & Tools

AI Model Improves Trustworthy Flood Prediction with Explainability

Researchers developed Context-Aware Concept Distillation (CACD), a framework that distills opaque Deep Learning models into interpretable, hydrology-aware surrogates for flood prediction. This method provides verifiable causal narratives required by disaster response authorities, achieving high fidelity and outperforming black-box baselines globally.

Eli Levinkopf, Efrat Morin, Claudia V. GoldmanJul 28, 2026
AI ResearchAI Engineering & DevTools

Diffusion Models' Generative Quality Gets Comprehensive Theoretical Analysis

This research provides a unified theoretical framework for understanding the generalization and convergence of score-based diffusion models. It decomposes the total generative error into four interpretable components, quantifying how training data, discretization, and optimization affect sample fidelity.

Jinshu Huang, Yiming Jiang, Chunlin WuJul 28, 2026