Error Propagation in Multi-turn Dialogue

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2272 papers417 conceptsfrozen 2026-10-04

Large Language Models (LLMs) often demonstrate impressive reasoning in isolated, single-turn interactions. However, as a conversation progresses, their performance frequently degrades. This decline is rarely a sudden collapse; instead, it is driven by error propagation, a phenomenon where a minor mistake in an early turn "poisons" the dialogue history, leading to a cascade of increasingly severe failures in subsequent turns.

The Snowballing Effect

In multi-turn settings, the model’s previous outputs become part of its current input context. If a model generates a factual hallucination or a logical error in turn two, it treats that error as a grounded truth for turn three. Kedi Chen and colleagues, in their work on DiaHalu, describe this as "hallucination snowballing."

The core of the problem is that LLMs rarely self-correct. Instead of identifying an internal inconsistency, they tend to double down on previous errors to maintain conversational fluidity. As Weihong Zhong demonstrates with the MMHalSnowball benchmark, models will even ignore direct visual evidence if it contradicts a mistake they made earlier in the dialogue, prioritizing historical consistency over external reality.

Why Single-Turn Benchmarks Fail

Traditional evaluation metrics—like those focusing on per-response accuracy—often mask the fragility of these systems. A model might have a 90% success rate on individual turns, but in a 10-turn conversation, the probability of completing the entire task successfully can drop toward zero.

Research by Wai-Chung Kwan on MT-Eval highlights this gap. By replacing a model's erroneous intermediate responses with "gold" (correct) human answers, researchers found that models could often recover their reasoning capabilities. This proves that the failure isn't necessarily a lack of knowledge, but rather an inability to navigate the "noisy" context created by their own previous mistakes. Similarly, Souvik Das and the creators of the FADE dataset show that these cascading errors are particularly difficult to detect because the resulting dialogue remains internally coherent even as it drifts away from the truth.

Domain-Specific Fragility

The impact of error propagation is most visible in tasks requiring strict logical dependencies, such as coding or database querying:

  • Stepwise Refinement: In SR-Eval, researchers found that completion rates for complex code tasks drop to ~20% in multi-turn scenarios because a single bug in an early requirement refinement prevents all subsequent steps from succeeding.
  • Conversational SQL: Systems like RAT-SQL-TC struggle as conversations lengthen; a misunderstanding of a single table column in turn one can render every subsequent SQL query in the session invalid.
  • Goal-Oriented Tasks: In games like GuessWhat?!, a single incorrect "yes/no" answer early in a sequence of questions causes the entire search trajectory to fail, as the agent's mental model of the world becomes fundamentally misaligned with the environment.

Strategies for Mitigation

To address these compounding failures, researchers are moving toward more dynamic evaluation and planning frameworks. Yunan Lu and the team behind SAGE propose using "top-down" and "bottom-up" user simulators to stress-test agents, identifying long-tail defects that only emerge after several turns of interaction.

On the architectural side, some are moving away from purely sequential generation. Hanwen Du introduced DiffTOD, which uses diffusion models to plan entire dialogue trajectories at once. By considering the future and the past simultaneously, these models can maintain global consistency and avoid the "greedy" local errors that typically trigger a snowball effect. Other approaches, such as the "literalization" strategy proposed by Harsh Jhamtani, focus on pre-processing linguistic noise (like metaphors) to prevent the initial "poisoning" of the dialogue state.

Papers in this version

Showing 60 of 272 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 234 papers
  2. Version 1 · 2026-10-04 · 272 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_sequential_error_propagation_cascading_e_70d6c9a6,
  title = {sequential error propagation (cascading error propagation)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sequential-error-propagation-cascading-error-propagation/rdr_70d6c9a692e04536942cb2b94cda4d8b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_error_propagation_in_multi_turn_dialogue_70d6c9a6_v1,
  title = {Error Propagation in Multi-turn Dialogue},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/error-propagation-in-multi-turn-dialogue/rdr_70d6c9a692e04536942cb2b94cda4d8b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}