Skip to content
OpenSmartRoute

LLMs1 min read

LLM History Effects: Causal Degradation in Multi-Turn Interactions

Research reveals that LLM performance degrades across multi-turn interactions, and editing assistant-generated history can significantly impact results. This study explores the selective nature of these effects, offering insights for managing model context.

By OpenSmartRoute editorial · written through the router by writer-small

From arXiv cs.CL - “What if LLMs Ate Their Words: Causal History Effects in Multi-Turn Interaction”

The research investigated the impact of multi-turn interaction on LLM performance. Experiments compared fully specified single-turn input (FULL) with progressively revealed multi-turn interaction (SHARDED) across six task families and five models. Degradation was observed and was dependent on both the task and the model. The study then analyzed completed SHARDED conversations by replaying user messages while editing assistant-generated history – a process termed neutralization. This resulted in a +.027 change in min-max normalized performance across 2,973 trajectories.

Further investigation, termed Turn Surgery, involved altering individual assistant turns. Analysis of 237 degraded trajectories showed that 63.7% contained at least one beneficial intervention, but most positions remained unchanged. In binary tasks, 48.4% experienced a fail-to-success reversal with a single intervention. The research found that task-dependent rather than universal internal signatures linked behaviorally consequential history changes to downstream state differences.

Overall, the research demonstrated that assistant-generated history has active but selective effects on multi-turn performance. This suggests a need for selective rather than uniform history management strategies. The study highlights the importance of understanding how model context influences behavior, particularly in extended interactions.

Source: https://arxiv.org/abs/2609.05882v1

Published Sep 9, 2026 · updated Sep 9, 2026 · 182 words

Keep reading

Related posts

More in LLMs

LLMs1 min read

Query Rewriting Enhances RAG Performance Through Complementarity

Research found that combining query rewriting strategies with a strong RAG baseline yields significant improvements in retrieval accuracy, primarily driven by complementarity rather than increased retrieval budget. A cost-aware router further optimizes this approach.

LLMs1 min read

Intra-Prompt Parallel Decoding for CCQA

A new inference method, Intra-Prompt Parallel Decoding (IPPD), achieves up to 7x throughput in common-context question answering by decoding multiple questions within a single prompt. This approach overcomes GPU memory bottlenecks and outperforms existing techniques like prefix caching.

LLMs1 min read

ModularPhaseNet: Discretized Phase Geometry for Transformers

ModularPhaseNet introduces a classical, integer-computable discretization of complex phase geometry, offering a new approach to Transformer architecture. It utilizes cyclic group operations for phase representation, aiming to improve semantic hierarchy and contextual consistency.

LLM History Effects: Causal Degradation in Multi-Turn Interactions - OpenSmartRoute