Conversation history degrades LLM reasoning accuracy by up to 29 points
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers have quantified a critical degradation pattern in multi-turn LLM systems: as conversation history accumulates, model accuracy drops sharply even when task difficulty remains constant. GPT-4o-mini and Mistral-7B showed 21-29 percentage point accuracy losses in math reasoning tasks, while a new benchmark called CDR-Stream revealed accuracy collapsing from 60% to 10% within twenty dialogue turns. This finding challenges assumptions about LLM reasoning robustness in production systems and suggests that process-level quality metrics may not reliably warn before performance failure. The work has immediate implications for deployed multi-turn assistants and raises questions about whether current architectural approaches to context management are fundamentally limited.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The paper isolates a specific failure mode: accuracy decay tied purely to conversation length, not task complexity. This is distinct from prior work showing multi-turn brittleness because it demonstrates the degradation happens even when individual turns remain equally difficult.
This finding extends a pattern documented across recent benchmarks. ToolTurns (November 11) showed multi-turn performance collapse from 70% to 40% in agent tool use, and the Critic paper (November 14) revealed how iterative feedback compounds errors across turns. Both flagged multi-turn fragility as a deployment risk. This new work quantifies the mechanism: it's not just that models fail on harder problems across turns, but that context accumulation itself degrades reasoning capacity. The implication is sharper: even static task difficulty cannot protect production systems from accuracy loss over conversation length.
If the CDR-Stream benchmark is adopted by major model providers for pre-release testing and leads to architectural changes (context pruning, sliding windows, or retrieval-augmented history) before Q2 2027, that signals the community treats this as a solvable problem worth engineering resources. If models released after this paper show no improvement on CDR-Stream's 20-turn collapse metric, it indicates the finding is real but not prioritized.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
Multi-turn dialogue exposes 30-point performance cliff in LLM agents
Researchers have exposed a critical gap in how language models handle real-world tool use. ToolTurns reveals that while leading agents succeed on isolated requests 70% of the time, performance collapses to 40% when tasks span multiple conversational turns requiring context retention and clarification. This benchmark of 2,620 problems across 1,827 tools signals that production deployments…
MentionsGPT-4o-mini · Mistral-7B · Llama-4 Scout · CDR-Stream
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “The Silent Cost of Conversation History in LLM Reasoning”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.