Researchers Identify 'Contextual Inertia' Bug in LLMs During Multi-Turn Conversations
Researchers have identified a critical failure mode in large language models called 'contextual inertia'—where models ignore new information in multi-turn conversations and rigidly stick to previous reasoning. A new training method called RLSTA uses single-turn performance as an anchor to stabilize multi-turn reasoning and recover performance lost to this phenomenon.
Researchers Identify 'Contextual Inertia' Bug in LLMs During Multi-Turn Conversations
A new arXiv paper identifies a fundamental instability in how large language models handle multi-turn interactions: they systematically ignore new information provided by users and double down on earlier (often incorrect) reasoning paths.
The researchers term this failure mode contextual inertia—a phenomenon where models maintain consistency with previous outputs even when users explicitly provide corrections, constraints, or updated data in subsequent turns. This directly contradicts the models' strong single-turn performance on the same problems.
The Problem: Stark Performance Collapse
When information is revealed incrementally or requires updates across turns, LLMs exhibit "substantial vulnerability" compared to their baseline single-turn performance. The paper demonstrates that models fail to integrate new constraints and often collapse entirely in performance—despite having the raw capability to solve the problem if all information were provided upfront.
This isn't a reasoning deficit. It's a behavioral rigidity: models prioritize consistency with their prior outputs over adapting to fresh input.
The Solution: RLSTA
Researchers propose Reinforcement Learning with Single-Turn Anchors (RLSTA), a generalizable training approach that leverages a model's superior single-turn capabilities as internal reward signals. The method works by:
- Using single-turn performance as a stable "anchor" to evaluate multi-turn behavior
- Training models via reinforcement learning to align multi-turn responses with these anchors
- Enabling models to self-calibrate reasoning based on the latest information
Key Findings
- RLSTA significantly outperforms standard fine-tuning and abstention-based baselines
- The method shows strong cross-domain generalization (performance transfers from math to code problems, for example)
- Effectiveness holds even without external verifiers—important for general-domain applicability
- Models trained with RLSTA actively break contextual inertia and adapt to new constraints
What This Means
Contextual inertia appears to be a widespread vulnerability in current LLMs that existing training methods don't address. Multi-turn interaction is increasingly central to real-world AI applications—from iterative debugging with code assistants to step-by-step problem-solving with constraints that evolve. This work identifies a concrete mechanism behind multi-turn failure and provides a practical fix that doesn't require domain-specific verifiers. The cross-domain generalization results suggest RLSTA could become a standard training component for more robust LLM systems, though adoption will depend on the computational cost of the reinforcement learning phase and whether it scales to very large models.