Lost in the Noise: How Reasoning Models Fail with Contextual Distractors

TL;DR

This work exposes a substantial gap between clean benchmark performance and real-world robustness by introducing NoisyBench, a suite that stress-tests reasoning models under contextual distractors. It demonstrates dramatic performance degradation, especially with hard negatives, and shows that agentic workflows can amplify these errors. To address this, the authors propose NoisyInstruct for exposure-based robustness training and a Rationale-Aware Reward (RARE) that incentivizes grounding reasoning in useful information within noise, yielding substantial robustness gains when combined with reinforcement learning. The study also reveals structural patterns, such as inverse scaling with increased test-time reasoning and attention biases toward distractors, offering actionable guidance for building resilient, reasoning-capable agent systems.

Abstract

Recent advances in reasoning models and agentic AI systems have led to an increased reliance on diverse external information. However, this shift introduces input contexts that are inherently noisy, a reality that current sanitized benchmarks fail to capture. We introduce NoisyBench, a comprehensive benchmark that systematically evaluates model robustness across 11 datasets in RAG, reasoning, alignment, and tool-use tasks against diverse noise types, including random documents, irrelevant chat histories, and hard negative distractors. Our evaluation reveals a catastrophic performance drop of up to 80% in state-of-the-art models when faced with contextual distractors. Crucially, we find that agentic workflows often amplify these errors by over-trusting noisy tool outputs, and distractors can trigger emergent misalignment even without adversarial intent. We find that prompting, context engineering, SFT, and outcome-reward only RL fail to ensure robustness; in contrast, our proposed Rationale-Aware Reward (RARE) significantly strengthens resilience by incentivizing the identification of helpful information within noise. Finally, we uncover an inverse scaling trend where increased test-time computation leads to worse performance in noisy settings and demonstrate via attention visualization that models disproportionately focus on distractor tokens, providing vital insights for building the next generation of robust, reasoning-capable agents.

Figures (45)

• Figure 1: Comparison between clean benchmarks and NoisyBench, showing that models perform well in sterilized settings but fail under realistic noise from random documents, irrelevant chat history, and hard negative distractors, which reveals weaknesses in alignment, reasoning, and RAG robustness.
• Figure 2: Agentic Workflow Results. Agentic workflows improve performance in the clean setting (ND) but degrade under noisy conditions (RD, RC, HN).
• Figure 3: Context Engineering Results. Context engineering methods (GEPA, DC, ACE) show limited gains over the base model under noisy distractors (HN, RD, RC).
• Figure 4: Reward Dynamics during RL Training. RARE steadily reduces distracted chains of thought while increasing outcome-based rewards, which leads to higher final accuracy compared to training with outcome-only rewards (OR).
• Figure 5: As distractor similarity increases across benchmarks, accuracy consistently decreases while average reasoning token usage increases.