# Graph-based Target Back-Propagation for Context Adaptation in Multi-LLM Agentic Systems

Source: [Graph-based Target Back-Propagation for Context Adaptation in Multi-LLM Agentic Systems](<https://www.sota2.com/research/paper/6a2f86ded33dfb23e783be04>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-06-16T05:00:05.515Z

Authors: Tan Zhu, Tong Yao, Kananart Kuwaranancharoen, Amit Singh, Yushang Lai, Deepa Mohan, Shankara Bhargava

Published: 2026-06-12T06:27:15.000Z

## Abstract

Context adaptation automates prompt engineering in LLM-based systems by iteratively revising tunable prompts from task feedback, without modifying model weights. Extending this paradigm to multi-LLM agentic systems is crucial: existing methods suffer from inaccurate credit assignment and lack convergence guarantees. We propose \\textbf{G}raph-based \\textbf{T}arget \\textbf{B}ack-\\textbf{P}ropagation (GTBP), a context adaptation framework for agentic workflows modeled as directed acyclic graphs. GTBP propagates local target outputs backward through the workflow graph and uses target--output discrepancies to guide a stage-wise prompt update mechanism. Theoretically, we show that GTBP's stage-wise prompt updates become stable over iterations, and that a sufficiently capable LLM optimizer can decrease the overall objective. Empirically, GTBP consistently outperforms strong baselines across three benchmarks while maintaining comparable computational cost.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Mathematical Reasoning on LiveBench-Math July 30, 2025 snapshot (test)](<https://www.sota2.com/research/sota/mathematical-reasoning-on-livebench-math-july-30-2025-snapshot-test.md>): Mathematical Reasoning; LiveBench-Math July 30, 2025 snapshot (test); Accuracy: 66.9.

- [Multi-hop Question Answering on HotpotQA (val) (F1)](<https://www.sota2.com/research/sota/multi-hop-question-answering-on-hotpotqa-val-f1.md>): Multi-hop Question Answering; HotpotQA (val); F1 Score: 80.

- [Population Subgroup Response Prediction on SubPOP (test)](<https://www.sota2.com/research/sota/population-subgroup-response-prediction-on-subpop-test.md>): Population Subgroup Response Prediction; SubPOP (test); WD: 0.1072.

- [Demographic Bias Assessment on SubPOP](<https://www.sota2.com/research/sota/demographic-bias-assessment-on-subpop.md>): Demographic Bias Assessment; SubPOP; Error Rate (Northeast): 0.0952.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2606.14155>)

[PDF](<https://arxiv.org/pdf/2606.14155.pdf>)