Public benchmarks enable fair and reproducible evaluation of LLM reasoning, but they become fragile for deep research agents that actively search the web during inference. Such agents may retrieve public benchmark metadata, question context, or even ground-truth answers via web search. This gives rise to Search-Time Contamination (STC), where external retrieval bypasses intended reasoning and inflates measured performance. We systematically study STC in deep research agent evaluation. We define three contamination types with increasing severity, namely Benchmark Metadata Leakage, Question-Context Leakage, and Explicit Answer Leakage, and develop detection algorithms to identify them and quantify their impact on agent performance. Evaluating modern deep research agents on seven public benchmarks, we find that STC is widespread and can inflate performance by up to 4%. Our findings show that existing evaluations may overestimate true reasoning ability. We therefore advocate contamination-aware practices, including isolated sandboxes, transparent search trajectories, and controlled benchmark access.

Speculative decoding accelerates large language model (LLM) inference by allowing a lightweight draft model to propose candidate tokens for verification by a target model. In real-world deployments, however, shifts in input distributions and updates to the target model can weaken draft–target alignment, reducing token acceptance and acceleration. Repeatedly retraining the draft model is costly, delays deployment, and may impair performance on previously encountered distributions. Achieving online alignment therefore requires rapid and stable adaptation while incorporating real-time feedback within a limited compute budget. To this end, we propose Dynamic Speculative Decoding (DySD), a framework that aligns the draft model with the target model during inference without offline retraining. DySD combines hierarchical feature alignment with distribution-level matching and uses divergence-triggered dynamic distillation to selectively update the draft model. Experiments across multiple datasets show that DySD achieves 1.6×–2.0× speedups and boosts token acceptance rates by 10%–30% under distribution shifts, enabling speculative decoding to remain fast, stable, and adaptive in ever-changing environments.

Natural language counterfactual generation aims to minimally modify a given text such that the modified text will be classified into a different class. The generated counterfactuals provide insight into the reasoning behind a model’s predictions by highlighting which words significantly influence the outcomes. Additionally, they can be used to detect model fairness issues and augment the training data to enhance the model’s robustness. A substantial amount of research has been conducted to generate counterfactuals for various NLP tasks, employing different models and methodologies. With the rapid growth of studies in this field, a systematic review is crucial to guide future researchers and developers. To bridge this gap, this survey provides a comprehensive overview of textual counterfactual generation methods, particularly those based on Large Language Models. We propose a new taxonomy that systematically categorizes the generation methods into four groups and summarizes the metrics for evaluating the generation quality. Finally, we discuss ongoing research challenges and outline promising directions for future work.

To encourage fairness and transparency, there exists an urgent demand for deriving reliable explanations for large language models (LLMs). One promising solution is concept-based explanations, i.e., human-understandable concepts from internal representations. However, due to the compositional nature of languages, current methods mostly discover correlational explanations instead of causal features. Therefore, we propose a novel framework to provide impact-aware explanations for users to understand the LLM’s behavior, which are robust to feature changes and influential to the model’s predictions. Specifically, we extract predictive high-level features (concepts) from the model’s hidden layer activations. Then, we innovatively optimize for features whose existence causes the output predictions to change substantially. Extensive experiments on real and synthetic tasks demonstrate that our method achieves superior results on predictive impact, explainability, and faithfulness compared to the baselines, especially for LLMs.