Large language models excel as few-shot learners when provided with appropriate demonstrations, yet this strength becomes problematic in multi-turn agent scenarios, where LLMs erroneously mimic their own previous responses as few-shot examples. Through attention analysis, we identify conversational inertia, a phenomenon where models exhibit strong diagonal attention to previous responses, which is associated with imitation bias that constrains exploration. This reveals a tension when transforming few-shot LLMs into agents: longer context enriches environmental feedback for exploitation, yet also amplifies conversational inertia that undermines exploration. Our key insight is that for identical states, actions generated with longer contexts exhibit stronger inertia than those with shorter contexts, enabling construction of preference pairs without environment rewards. Based on this, we propose Context Preference Learning to calibrate model preferences to favor low-inertia responses over high-inertia ones. We further provide context management strategies at inference time to balance exploration and exploitation. Experimental results across eight agentic environments and one deep research scenario validate that our framework reduces conversational inertia and achieves performance improvements.
Multi-turn LLM agents solve tasks by repeatedly observing the environment, generating actions, and appending the resulting interaction history to the next prompt. This design gives the model useful feedback for exploitation, but it also changes the role of context: previous assistant outputs can start to look like few-shot demonstrations written by the model itself.
We find that longer histories do not always help. Agent performance can degrade at moderate context lengths that remain far below modern model capacity limits. Attention analysis suggests a concrete mechanism behind this degradation: models increasingly attend to their own previous responses, especially along token-aligned diagonal patterns, and become biased toward repeating earlier behavior.
We call this failure mode conversational inertia. It appears as strong diagonal attention to previous assistant responses and can make the model reproduce the structure of earlier actions instead of adapting to new observations. This motivates our two complementary mitigations: Context Preference Learning and Clip Context.
Our framework targets inertia at two levels. At the model level, Context Preference Learning teaches the model to prefer lower-inertia actions generated from shorter contexts. At the inference level, Clip Context periodically clears older interaction history to reset inertial patterns while retaining enough recent state for exploitation.
CPL constructs long-short context preference pairs from the same state. Actions generated with shorter contexts tend to exhibit weaker inertia, so they serve as preferred responses over actions generated from longer histories.
The preference signal is reward-free: it does not require ground-truth task success, human annotation, or environment-specific reward modeling.
Clip Context clears older interaction turns at fixed intervals, creating cycles of exploration and exploitation. The agent can use recent context, but accumulated self-imitation is periodically broken.
Unlike sliding windows, clipping preserves long append-only segments between resets, which makes the method naturally compatible with KV cache reuse.
We evaluate across eight agent environments that cover navigation, web interaction, text games, reasoning-dependent strategic games, and crafting scenarios. This makes the setting broad enough to test both exploration and exploitation under multi-turn interaction histories.
Across eight agent environments, Clip Context consistently improves over long-context and window-context baselines. Combined with CPL, the average score reaches 72.5%, compared with 68.9% for base Clip Context.
| Model | Variant | Context | MZ | ALF | WS | TC | FL | HM | 2048 | RH | Avg |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3-8B | Base | Window | 74.0 | 68.2 | 40.0 | 71.3 | 67.5 | 85.3 | 66.9 | 45.8 | 64.9 |
| Long | 39.0 | 61.0 | 23.3 | 67.5 | 63.5 | 75.1 | 65.6 | 40.3 | 54.4 | ||
| Summarization | 78.0 | 71.0 | 46.5 | 80.5 | 64.6 | 88.9 | 70.6 | 50.4 | 68.8 | ||
| Clip | 83.0 | 67.7 | 44.4 | 82.3 | 67.5 | 85.1 | 70.9 | 50.2 | 68.9 | ||
| CPL | Window | 78.0 | 67.5 | 44.9 | 74.5 | 68.2 | 89.5 | 72.4 | 44.3 | 67.4 | |
| Long | 44.0 | 59.5 | 40.5 | 65.5 | 70.0 | 77.7 | 70.2 | 44.8 | 59.0 | ||
| Summarization | 87.5 | 71.5 | 52.2 | 75.2 | 69.4 | 89.9 | 71.6 | 49.1 | 70.8 | ||
| Clip | 91.5 | 70.3 | 54.9 | 83.0 | 68.4 | 87.5 | 73.0 | 51.1 | 72.5 | ||
| Llama3.1-8B | Base | Window | 79.7 | 21.5 | 45.4 | 68.2 | 58.6 | 87.0 | 70.1 | 38.0 | 58.6 |
| Long | 48.0 | 20.3 | 29.1 | 49.1 | 54.0 | 79.9 | 68.1 | 38.2 | 48.3 | ||
| Summarization | 81.5 | 31.5 | 37.6 | 62.7 | 53.8 | 88.3 | 69.8 | 38.0 | 57.9 | ||
| Clip | 82.5 | 28.5 | 43.6 | 70.0 | 55.9 | 84.1 | 68.7 | 40.5 | 59.2 | ||
| CPL | Window | 79.0 | 25.3 | 44.9 | 72.0 | 55.4 | 86.5 | 67.8 | 38.9 | 58.7 | |
| Long | 46.5 | 18.8 | 32.3 | 46.5 | 52.4 | 78.5 | 69.7 | 38.8 | 47.9 | ||
| Summarization | 77.2 | 31.5 | 39.5 | 65.0 | 56.5 | 88.6 | 68.0 | 40.4 | 58.3 | ||
| Clip | 82.3 | 31.5 | 43.8 | 70.0 | 56.8 | 86.3 | 68.6 | 39.3 | 59.8 | ||
| GPT-4o-mini | / | Window | 82.0 | 52.5 | 37.8 | 55.0 | 89.4 | 97.3 | 73.4 | 40.2 | 66.0 |
| Long | 62.0 | 62.0 | 45.5 | 52.0 | 87.2 | 96.8 | 72.9 | 38.4 | 64.6 | ||
| Summarization | 71.0 | 56.2 | 42.2 | 66.7 | 91.2 | 98.9 | 72.9 | 40.7 | 67.5 | ||
| Clip | 81.0 | 54.5 | 39.8 | 81.8 | 92.0 | 98.9 | 75.5 | 45.3 | 71.1 |
We also evaluate long-term reasoning on BrowseComp, where the agent must balance information gathering, search depth, and the risk of answering too early. Clip-12to6 obtains the best score among the tested settings and remains competitive with summarization while avoiding extra summary-generation overhead.
| Method | Score | Proactive Rate | Proactive Score |
|---|---|---|---|
| Window-6 | 25.0 ± 1.7 | 27.3% | 55.7 |
| Window-9 | 23.4 ± 1.5 | 26.6% | 58.8 |
| Window-12 | 24.2 ± 1.8 | 25.4% | 58.5 |
| Clip-12to0 | 25.0 ± 1.8 | 29.3% | 56.0 |
| Sum-12to0 | 27.7 ± 2.0 | 63.7% | 37.4 |
| Clip-12to6 | 29.3 ± 1.5 | 30.9% | 61.0 |
| Sum-12to6 | 28.1 ± 1.6 | 30.8% | 63.3 |
The two mitigation mechanisms affect attention differently. Clip Context directly changes the available history and reduces diagonal attention by removing accumulated assistant traces. Context Preference Learning reduces diagonal attention while keeping overall assistant attention comparatively stable, suggesting that the model learns to use history less positionally and less imitatively.
Clip Context is not only a behavioral intervention. Because it preserves append-only generation between clip operations, it can reuse KV cache more effectively than sliding-window truncation. In our speed benchmark, Clip achieves a 2-7x prefill speedup over window methods while maintaining strong task performance.
@article{wan2026mitigating,
author = {Wan, Yang and Cao, Zheng and Zhang, Zhenhao and Zeng, Zhengwen and Shen, Shuheng and Meng, Changhua and Zhu, Linchao},
title = {Mitigating Conversational Inertia in Multi-Turn Agents},
journal = {arXiv preprint arXiv:2602.03664},
year = {2026},
}