Early Stopping for Large Reasoning Models via Confidence Dynamics

Large reasoning models rely on long chain-of-thought generation to solve complex problems, but extended reasoning often incurs substantial computational cost and can even degrade performance due to overthinking. A key challenge is determining when the model should stop reasoning and produce the final answer. In this work, we study the confidence of intermediate answers during reasoning and observe two characteristic behaviors: correct reasoning trajectories often reach high-confidence answers early, while incorrect rollouts tend to produce long, unproductive reasoning traces and exhibit less reliable confidence dynamics. Motivated by these observations, we propose CoDE-Stop (Confidence Dynamics Early Stop), an early stopping method that leverages the dynamics of intermediate answer confidence to decide when to terminate reasoning, requiring no additional training and easily integrating into existing models. We evaluate CoDE-Stop on diverse reasoning and science benchmarks across multiple models. Compared to prior early stopping methods, it achieves a more favorable accuracy-compute tradeoff and reduces total token usage by 25-50% compared to standard full-length reasoning. In addition, we provide analyses of confidence dynamics during reasoning, offering insights into how confidence changes in both correct and incorrect trajectories.

Gaming Tool Preferences in Agentic LLMs

Large language models (LLMs) can now access a wide range of external tools, thanks to the Model Context Protocol (MCP). This greatly expands their abilities as various agents. However, LLMs rely entirely on the text descriptions of tools to decide which ones to use--a process that is surprisingly fragile. In this work, we expose a vulnerability in prevalent tool/function-calling protocols by investigating a series of edits to tool descriptions, some of which can drastically increase a tool's usage from LLMs when competing with alternatives. Through controlled experiments, we show that tools with properly edited descriptions receive over 10 times more usage from GPT-4.1 and Qwen2.5-7B than tools with original descriptions. We further evaluate how various edits to tool descriptions perform when competing directly with one another and how these trends generalize or differ across a broader set of 10 different models. These phenomenons, while giving developers a powerful way to promote their tools, underscore the need for a more reliable foundation for agentic LLMs to select and utilize tools and resources.

Chain-of-Defensive-Thought: Structured Reasoning Elicits Robustness in Large Language Models against Reference Corruption

Chain-of-thought prompting has demonstrated great success in facilitating the reasoning abilities of large language models. In this work, we explore how these enhanced reasoning abilities can be exploited to improve the robustness of large language models in tasks that are not necessarily reasoning-focused. In particular, we show how a wide range of large language models exhibit significantly improved robustness against reference corruption using a simple method called chain-of-defensive-thought, where only a few exemplars with structured and defensive reasoning are provided as demonstrations. Empirically, the improvements can be astounding, especially given the simplicity and applicability of the method. For example, in the Natural Questions task, the accuracy of GPT-4o degrades from 60% to as low as 3% with standard prompting when 1 out of 10 references provided is corrupted with prompt injection attacks. In contrast, GPT-4o using chain-of-defensive-thought prompting maintains an accuracy of 50%.

Seeing What's Not There: Spurious Correlation in Multimodal LLMs

Unimodal vision models are known to rely on spurious correlations, but it remains unclear to what extent Multimodal Large Language Models (MLLMs) exhibit similar biases despite language supervision. In this paper, we investigate spurious bias in MLLMs and introduce SpurLens, a pipeline that leverages GPT-4 and open-set object detectors to automatically identify spurious visual cues without human supervision. Our findings reveal that spurious correlations cause two major failure modes in MLLMs: (1) over-reliance on spurious cues for object recognition, where removing these cues reduces accuracy, and (2) object hallucination, where spurious cues amplify the hallucination by over 10x. We validate our findings in various MLLMs and datasets. Beyond diagnosing these failures, we explore potential mitigation strategies, such as prompt ensembling and reasoning-based prompting, and conduct ablation studies to examine the root causes of spurious bias in MLLMs. By exposing the persistence of spurious correlations, our study calls for more rigorous evaluation methods and mitigation strategies to enhance the reliability of MLLMs.

Decompose-and-Compose: A Compositional Approach to Mitigating Spurious Correlation

While standard Empirical Risk Minimization (ERM) training is proven effective for image classification on in-distribution data, it fails to perform well on out-of-distribution samples. One of the main sources of distribution shift for image classification is the compositional nature of images. Specifically, in addition to the main object or component(s) determining the label, some other image components usually exist, which may lead to the shift of input distribution between train and test environments. More importantly, these components may have spurious correlations with the label. To address this issue, we propose Decompose-and-Compose (DaC), which improves robustness to correlation shift by a compositional approach based on combining elements of images. Based on our observations, models trained with ERM usually highly attend to either the causal components or the components having a high spurious correlation with the label (especially in datapoints on which models have a high confidence). In fact, according to the amount of spurious correlation and the easiness of classification based on the causal or non-causal components, the model usually attends to one of these more (on samples with high confidence). Following this, we first try to identify the causal components of images using class activation maps of models trained with ERM. Afterward, we intervene on images by combining them and retraining the model on the augmented data, including the counterfactual ones. Along with its high interpretability, this work proposes a group-balancing method by intervening on images without requiring group labels or information regarding the spurious features during training. The method has an overall better worst group accuracy compared to previous methods with the same amount of supervision on the group labels in correlation shift.