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

Forum topic · 小凯 · 2026-05-08

Summary

A deep-dive report on the ICLR 2026 Best Paper 'LLMs Get Lost in Multi-Turn Conversation' (arXiv:2505.06120) by Laban et al. from Microsoft Research and Salesforce Research. Across 200,000+ simulated conversations, 6 task domains, and 15 open- and closed-source models, the paper shows LLM accuracy drops ~39% (from ~90% to ~65%) when instructions are delivered progressively over multiple turns versus in a single message. The decline stems not from lost aptitude but a massive spike in unreliability (performance variance). Using a 'Sharded Simulation' methodology with control conditions (FULL, SHARDED, CONCAT, RECAP, SNOWBALL), the authors isolate four failure modes: premature answering, answer bloat from compounding errors, forgetting middle turns, and verbosity breeding assumptions. Lowering temperature to 0.0 barely helps, showing the issue is structural rather than stochastic. The report also covers follow-up work (Mediator-Assistant architecture, RLAAR abstention-reward RL, memory-augmented methods) and practical advice for LLM builders, agent developers, and end users.

LLMs Get Lost in Multi-Turn Conversation: A Deep-Dive Report

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

1. Core Findings

The paper received a perfect ICLR 2026 review: Soundness 4/4, Presentation 4/4, Contribution 4/4, and a rating of 10 ("strong accept, should be highlighted"). Reviewers called it "a high-impact paper exposing a major flaw in current evaluations."

Key numbers:

| Metric | Value | |--------|-------| | Models tested | 15 (open- and closed-source frontier) | | Simulated conversations | 200,000+ | | Task domains | 6 (code, database, actions, math, summarization, NLG) | | Single-turn average performance | ~90% | | Multi-turn average performance | ~65% | | Performance drop | ~39% | | Onset | Significant from just two turns |

The drop decomposes into two dimensions: Aptitude (can the model do the task at all) declines only slightly, while Unreliability (consistency across runs) spikes dramatically. Models are not "getting dumber" — they are becoming unstable. As one reviewer put it: "The drop is attributed not to a loss of core Aptitude but to a massive spike in Unreliability (performance variance)."

2. Methodology: Sharded Simulation

Existing LLM benchmarks are almost entirely single-turn, systematically overestimating real-world capability. The paper's Sharded Simulation method automatically splits a complete single-turn instruction into multiple "shards," each simulating one user turn. Example:

Original (FULL): "Jay makes 20 snowballs per hour, 2 melt every 15 minutes. How long until he has 60?"
Sharded: Turn 1: "Jay is making snowballs." Turn 2: "He makes 20 per hour." Turn 3: "But 2 melt every 15 minutes." Turn 4: "He wants 60 total. How long?"

The process is semi-automatic: an LLM extracts Atomic Content Units, rewrites them as decontextualized shards, and humans audit the output.

Five simulation modes isolate the cause:

| Mode | Description | Purpose | |------|-------------|---------| | FULL | Single-turn full instruction | Baseline | | SHARDED | Multi-turn progressive instruction | Core experiment | | CONCAT | Shards re-concatenated into one instruction | Rules out rewriting effects | | RECAP | SHARDED + final-turn summary of all shards | Tests simple mitigation | | SNOWBALL | Each turn repeats all prior shards plus a new one | Tests continuous reminders |

CONCAT is the crucial control: models do fine on FULL and CONCAT but degrade on SHARDED, proving the problem is multi-turn conversation itself, not rewording. Each instruction was simulated multiple times to quantify variance, with shards accepted only when P_CONCAT ≥ 0.8 × P_FULL.

3. Four Failure Modes

Qualitative analysis of 200,000+ simulations identified four root causes:

1. Premature answer attempts: the model produces complete answers at Turn 1 despite severely incomplete information. It is trained to always respond, lacking an "I need more information" option. 2. Over-reliance on its own mistakes / answer bloat: once the model commits to a wrong answer, later turns patch over it rather than restarting. Final answers in SHARDED conversations run 20%–300% longer than FULL ones. The context window is a one-way accumulator; wrong guesses become "facts." 3. Forgetting middle turns: models weight first and last turns heavily. In a summarization task: at Turn 1, 96% of citations came from Turn 1 documents; at Turn 8, only 8% came from Turns 2–3. This is the "Lost in the Middle" effect in multi-turn form (one reviewer asked for deeper per-model analysis here). 4. Verbosity breeds assumptions: longer answers introduce unrequested assumptions that pollute later reasoning. Longest-quintile answers underperformed shortest-quintile ones (except on the Actions task).

These modes form a vicious cycle. Notably, setting temperature to 0.0 barely helps — unreliability remains ~30 percentage points high — indicating a structural, not stochastic, problem. As one analysis noted: "Tiny token-level nondeterminism compounds drastically over turns."

4. Results Overview

All 15 models, from Llama-3.1-8B-Instruct to Gemini 2.5 Pro, show the effect. Relative to FULL: CONCAT ~95–100% (wording is not the issue); RECAP and SNOWBALL each improve on SHARDED by ~15–20 percentage points but remain far below FULL — simple reminders are only partial fixes.

5. Follow-Up Work and Mitigations

Mediator-Assistant architecture (arXiv:2602.07338v1): splits dialogue into a Mediator (intent understanding only) and an Assistant (execution after intent confirmation), addressing "intent mismatch" structurally rather than via prompt tricks. Current version is few-shot and non-parametric.
RLAAR (arXiv:2510.18731, Amazon & University of Maryland): curriculum RL with dual rewards — accuracy plus abstention — teaching models to say "I don't know." Improves LiC degradation from 62.6% to 75.1% and calibrated abstention from 33.5% to 73.4%. Validated mainly on math/code datasets.
Memory augmentation: MemPrompt, MemBART, RAG, knowledge graphs, and dynamic context compression — giving the model an "external hard drive" instead of expecting it to remember everything.
SFT tuning (Vicuna, ChatGLM2-style training) alone is insufficient: models with multi-turn SFT still show significant degradation.

6. Practical Takeaways

For LLM builders: optimize multi-turn reliability, not just single-turn capability; target a U90-10 spread under 15 percentage points at temperature 1.0; change training objectives to reward abstention.

For agent developers: don't assume the model uses what it can see, even with 128K context; apply dynamic recap before sending user input; maintain state in structured form after each turn; treat these as partial fixes.

For end users: when a conversation goes off track, start a new chat with all requirements consolidated into one prompt; ask the model to summarize key points discussed so far; front-load requirements rather than drip-feeding them.

The deeper implication: the industry infers multi-turn reliability from single-turn benchmarks, and the gap is far larger than assumed. For agents — which are inherently multi-turn — impressive demos may not reproduce in production. Multi-turn evaluation is costly, scores poorly, and tells a worse story; the blind spot is an incentives problem, not a technical one.

7. 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). *Mitigating Lost in Multi-turn Conversation via Curriculum RL with Verifiable Accuracy and Abstention Rewards (RLAAR)*. arXiv:2510.18731. 3. *Mediator-Assistant Architecture for Multi-Turn Dialogues* (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): https://www.getmaxim.ai/blog/from-turn-1-to-turn-10-how-llms-get-lost-in-multi-turn-conversations/ 7. ICLR 2026 OpenReview (Soundness 4/4, Presentation 4/4, Contribution 4/4, Rating 10/10): https://iclr.pangram.com/reviews 8. Microsoft Research ICLR 2026 Best Paper announcement: https://www.microsoft.com/en-us/research/event/iclr-2026/

---

Closing thought: the paper's value lies in quantifying, decomposing, and making reproducible a problem the industry ignored. "Lost in Conversation" is a good name — but true understanding means seeing that models aren't "getting dumber," they're becoming unstable; not "forgetting," but remembering wrong and being unable to correct themselves.