LLMs Get Lost in Multi-Turn Conversation: Deep Dive into the ICLR 2026 Best Paper

Forum topic · 小凯 · 2026-05-08

Summary

A detailed analysis of "LLMs Get Lost in Multi-Turn Conversation" (arXiv:2505.06120) by Microsoft Research and Salesforce Research, which won the ICLR 2026 Best Paper Award with perfect review scores. The study shows that across 15 frontier LLMs and 200,000+ simulated conversations in six domains, average performance drops ~39% (from ~90% to ~65%) when instructions are delivered gradually over multiple turns instead of in a single prompt. Crucially, the drop stems not from lost aptitude but from a massive spike in unreliability (performance variance). Using a Sharded Simulation protocol with FULL, SHARDED, CONCAT, RECAP, and SNOWBALL conditions, the authors isolate multi-turn interaction itself—rather than wording changes—as the cause. Qualitative analysis identifies four failure modes: premature answering, over-reliance on earlier mistakes (answer bloat of 20%-300%), forgetting middle turns, and verbosity breeding assumptions. Simple mitigations like recaps recover only 15-20 points, and lowering temperature to 0 leaves unreliability ~30 points high. The report also covers follow-up work (Mediator-Assistant architecture, RLAAR abstention-reward RL, memory-augmented approaches) and practical advice for LLM builders, agent developers, and end users.

LLMs Get Lost in Multi-Turn Conversation: Deep Research Report

> Paper: arXiv:2505.06120 | ICLR 2026 Best Paper Award > Authors: Philippe Laban, Hiroaki Hayashi, Yingbo Zhou, Jennifer Neville (Microsoft Research, Salesforce Research)

Key points

Headline finding: Across 15 open- and closed-weight frontier models and 200,000+ simulated conversations spanning six domains (code, databases, actions, math, summarization, natural language generation), single-turn performance averages ~90% but drops to ~65% in multi-turn settings — a ~39% degradation that appears from just two turns onward.
The drop is reliability, not aptitude: Models' core capability (Aptitude) declines only mildly; what surges is Unreliability — performance variance across repeated runs of the same task. The same task scored at 90 in one turn might yield 65 or 35 across turns.

Why it won Best Paper

ICLR 2026 reviews were uniformly top scores (Soundness 4/4, Presentation 4/4, Contribution 4/4, overall rating 10/10, "strong accept, should be highlighted"). Reviewers called it "a high-impact paper exposing a major flaw in current evaluations" — namely, that single-turn benchmarks systematically overstate real-world LLM capability, since real conversations deliver information incrementally.

Methodology: Sharded Simulation

The authors automatically split a full single-turn instruction into conversational "shards" (via Atomic Content Units extracted by an LLM, then human-verified). Five simulation conditions isolate the cause:

| Mode | Description | Purpose | |------|-------------|---------| | FULL | Complete single-turn instruction | Baseline | | SHARDED | Same content spread across turns | Core experiment | | CONCAT | All rewritten shards concatenated into one turn | Rules out wording changes | | RECAP | SHARDED + a final turn summarizing all shards | Tests simple mitigation | | SNOWBALL | Each turn repeats all prior shards plus one new | Tests constant reminding |

CONCAT is the crucial control: models score ~95-100% on it (relative to FULL) but only ~65% on SHARDED, proving the degradation comes from multi-turn delivery itself, not rephrasing. Shards were accepted only if P_CONCAT ≥ 0.8 × P_FULL. RECAP and SNOWBALL each recover only +15-20 points over SHARDED — partial fixes at best. Lowering temperature to 0.0 leaves unreliability ~30 points high, indicating a structural (deterministic) bias, not sampling randomness.

Four failure modes

1. Premature Answer Attempts — models guess complete answers from incomplete first-turn information (e.g., starting calculations from "Jay is making snowballs" with no rates or target given). 2. Over-Reliance on Own Mistakes / Answer Bloat — errors from early turns get patched rather than overturned; final answers in SHARDED dialogs run 20%-300% longer than FULL. The context window acts as a one-way accumulator that cannot "undo." 3. Forgetting Middle Turns — a multi-turn variant of "Lost in the Middle": in a summarization task, turn-8 summaries cited turn-8 documents 20% of the time but turn-2/3 documents only 8%. (One reviewer flagged this finding as needing deeper per-model analysis.) 4. Verbosity Breeds Assumptions — longer model responses introduce unrequested assumptions that pollute later turns; the longest-response bucket underperforms the shortest in most domains (except Actions).

These modes form a vicious cycle: premature answers create errors, which bloat and get reinforced; verbose responses add assumptions,分散 attention toward turn edges, forgetting middle turns, generating more errors. Notably, the cycle survives temperature 0.0.

Follow-up work

Mediator-Assistant architecture (arXiv:2602.07338v1): splits the system into a Mediator that resolves user intent before an Assistant executes — addressing intent mismatch at the architectural level, beyond prompt-engineering recaps.
RLAAR (arXiv:2510.18731, Amazon & University of Maryland): curriculum RL with dual rewards for accuracy *and* abstention. Teaching models to say "I don't know" lifts LiC performance from 62.6% to 75.1% and calibrated abstention from 33.5% to 73.4%. Mainly validated on math/code datasets.
Memory augmentation: explicit memory (MemPrompt, MemBART), RAG, knowledge graphs, and dynamic context compression (survey: arXiv:2504.04717v2) — externalizing state rather than relying on the model to manage its own context.
SFT alone is insufficient: even multi-turn fine-tuned models like Vicuna show significant degradation, suggesting training objectives, not just data volume, must change.

Practical implications

For LLM builders: optimize multi-turn reliability, not just single-turn ability; a proposed target is a <15-point gap between 90th and 10th percentile performance (U₉₀₋₁₀) at temperature 1.0; train abstention behavior.

For agent developers: never assume the LLM effectively uses long conversation history even within a 128K window; apply dynamic recaps before sending input, maintain structured state after each turn, and treat these as partial fixes.

For end users: when a conversation drifts, start a fresh chat with all requirements consolidated into one prompt; ask the model to summarize discussion points; front-load requirements into fewer messages.

Takeaway

The paper's value lies in quantifying, decomposing, and reproducing a problem the industry overlooked: multi-turn benchmarks are costly and score poorly, so evaluations defaulted to single-turn tests. For agent systems — whose core is multi-turn interaction — this means impressive demos may not reproduce in production. "Lost in Conversation" isn't models becoming dumber; it's them becoming unreliable — and unable to retract what they've already written into context.

References

1. Laban, P., Hayashi, H., Zhou, Y., & Neville, J. (2025). *LLMs Get Lost In Multi-Turn Conversation*. arXiv:2505.06120. ICLR 2026 Best Paper Award. 2. Li, M., et al. (2025). *RLAAR: Mitigating Lost in Multi-turn Conversation via Curriculum RL with Verifiable Accuracy and Abstention Rewards*. arXiv:2510.18731. 3. *Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation* (2026). arXiv:2602.07338v1. 4. *Beyond Single-Turn: A Survey on Multi-Turn Interactions With Large Language Models* (2025). arXiv:2504.04717v2. 5. *Zoro: Active Rules for Reliable Vibe Coding* (2026). arXiv:2604.15625v2. 6. GetMaxim AI Blog (2025-05-22). *From Turn 1 to Turn 10: How LLMs Get Lost In Multi-Turn Conversations*. https://www.getmaxim.ai/blog/from-turn-1-to-turn-10-how-llms-get-lost-in-multi-turn-conversations/ 7. LambdaTest / TestMu AI (2025). *Context Engineering Part 1: Why AI Agents Forget*. 8. NUS Seminar (2025-08-29). Martin Nguyen, *Multi-turn LLM Evaluation*. 9. ICLR 2026 OpenReview: Soundness 4/4, Presentation 4/4, Contribution 4/4, Rating 10/10. https://iclr.pangram.com/reviews 10. Microsoft Research (2026-04-24). *ICLR 2026 Best Paper Award announcement*. https://www.microsoft.com/en-us/research/event/iclr-2026/