Tool use is often assumed to monotonically improve reasoning, where external evidence is expected to help when relevant and be ignored when irrelevant. We show that this assumption fails in a state-dependent way. Across benchmarks with Python and Wikipedia tools, external evidence reliably helps when initial beliefs are weak, but can flip already-correct answers when those beliefs are strong. We frame this as a misallocation of revision authority, arguing that deferring to external evidence is suboptimal when internal support for the correct answer surpasses the tool’s expected output quality. This predicts that harm should concentrate on high-confidence no-tool cases. We test this prediction with threshold localization, wrong-trace audits, and a same-clue intervention showing that revision framing changes the damage caused by misleading evidence. These findings suggest that mixed no-tool/tool-assisted inference should arbitrate authority rather than privilege tool evidence by default. As a minimal demonstration, we introduce CASE, a label-free controller that selects between no-tool and tool-assisted trajectories using answer-state certainty and improves over existing confidence-based arbitration baselines.

Knowledge editing provides an efficient way to update factual knowledge in large language models. However, malicious edits may introduce safety risks, making it necessary to reverse undesirable editing effects. Existing reversal methods for parameter-modifying edits mainly focus on global removal, which may also erase beneficial edits that should be preserved. In this paper, we study selective reversal of edited knowledge, where the goal is to reverse targeted edited facts while preserving the remaining edited facts. Based on the hypothesis that each edit is sparsely encoded within the dominant subspace of the edited matrix, we propose a spectral-based reversal framework that locates edit-sensitive components within the dominant singular subspace of edited weights. Experiments across multiple settings demonstrate the effectiveness of our method in reversing selected edits while preserving unrelated edited facts. These results suggest that different edits are sparsely encoded within dominant singular components and can be separable when the number of edits is moderate, making selective spectral reversal a promising direction for locating edit-specific components and repairing edited language models.

Empowered by Large Language Models (LLMs), deep research systems integrate online information into research reports through multi-step web searching and information synthesis. Evaluating these reports presents a significant challenge due to their long context, high information density, and domain specificity, necessitating the use of AI evaluation as a substitute for prohibitively expensive human assessment. However, few studies have examined the effectiveness of existing reward models or LLMs in understanding and evaluating such research reports. To bridge this gap, we introduce DeepResearch RewardBench, the first benchmark dedicated to evaluating reward modeling for research reports. We collect diverse report candidates from multiple agents or tool-augmented LLMs. These candidates are scored using query-specific rubrics and judged LLMs, and we ensure data quality and difficulty through a combination of multi-round filtering and human verification. Our benchmark comprises preference datasets formatted as pairs, triplets, and quadruplets to test ranking capabilities across varying levels of difficulty. Extensive experiments demonstrate that while current reward models handle general coherence effectively, their performance significantly degrades when evaluating research reports. In contrast, generative LLMs acting as judges demonstrate comprehensive superiority in this domain, with comparative evaluation proving more effective than direct scoring methods. We anticipate that this benchmark will serve as a vital testbed for the evaluation and advancement of research agents.

Recently, the role of LLM-as-judge in evaluating large language models has gained prominence, emerging as an important method to partially replace costly human assessment. However, current judge models suffer from narrow specialization and limited robustness, undermining their capacity for comprehensive evaluations. In this work, we decompose the ability of a generalist and generative judge model into three levels: objective verification, subjective evaluation, and rubric refinement. We present CompassJudger-2, a novel generalist judge model that overcomes these limitations via a task-driven, multi-scenarios data curation strategy. We conducted large-scale data collection for each type of task and designed tailored rejection sampling strategies to filter the data, ensuring data diversity, accuracy, and effectiveness. Empirically, CompassJudger-2 achieves superior results across multiple judge and reward benchmarks, demostrating its excellent robustness and generalization ability. These contributions advance robust, scalable LLM judgment and establish new performance and evaluation standards.

Model merging is an effective technique for composing the capabilities of a multilingual model and a reasoning model. It has achieved promising generalization in multilingual reasoning tasks by aligning feature spaces of different models. However, the merged single model often fails to address the conflicts between source models, leading to suboptimal performance. In other words, the one-size-fits-all merging strategy may not align with the characteristics of different inputs which may require prioritizing certain models over others. To this end, we propose a Steerable Model Merging (ST-Merge) framework to modulate the contribution of each source model. To realize this idea, we introduce a gated cross-attention mechanism to weight or filter the two attended source models in an adaptive manner. Extensive experiments demonstrate that ST-Merge consistently outperforms multiple strong baselines on four multilingual reasoning benchmarks across 21 different languages.

Translation of natural language to first-order logical formula (NL-FOL) has recently gained significant attention for its critical role in logic-based NLP applications. Some studies attempt to utilize pretrained language models in a sequence-to-sequence manner for the NL-FOL task. However, these methods encounter challenges such as (1) inconsistency between the training and inference phases and (2) the data-intensive and resource-intensive finetuning process. This paper introduces a novel NL-FOL translation method, dubbed Code4Logic, which is based on in-context learning and employs code snippets to bridge the gap between natural language and first-order logic. By converting the translation task into a progressive code generation task, Code4Logic demonstrates strong generalization within a training-free manner, and enhances the performance of large language models (LLMs) to generate complex first-order logical formulas. Experimental results on NL-FOL task and downstream task datasets indicate that Code4Logic surpasses prominent training-free baselines and is comparable to supervised models trained on the full training data.

The semi-supervised learning (SSL) strategy in lightweight models requires reducing annotated samples and facilitating cost-effective inference. However, the constraint on model parameters, imposed by the scarcity of training labels, limits the SSL performance. In this paper, we introduce PS-NET, a novel framework tailored for semi-supervised text mining with lightweight models. PS-NET incorporates online distillation to train lightweight student models by imitating the Teacher model. It also integrates an ensemble of student peers that collaboratively instruct each other. Additionally, PS-NET implements a constant adversarial perturbation schema to further self-augmentation by progressive generalizing. Our PS-NET, equipped with a 2-layer distilled BERT, exhibits notable performance enhancements over SOTA lightweight SSL frameworks of FLiText and Disco in SSL text classification with extremely rare labelled data.

The rapid advancement of large language models (LLMs) has shown remarkable progress in complex reasoning tasks. However, a significant disparity exists between benchmark performances and real-world applications. We attribute this gap primarily to current evaluation protocols and metrics, which inadequately capture the full spectrum of LLM capabilities, especially in complex reasoning tasks where both accuracy and consistency are essential. In this paper, we introduce G-Pass@k, a novel evaluation metric that continuously assesses model performance across multiple sampling attempts, quantifying both the model’s performance potential and its stability. Through extensive experiments on various public and newly constructed benchmarks, we employ G-Pass@k in conjunction with state-of-the-art large language models to provide comprehensive insights into their potential capabilities and operational consistency. Our findings reveal a significant opportunity to enhance the realistic reasoning abilities of LLMs, underscoring the necessity for more robust evaluation metrics.

Answer verification is crucial not only for evaluating large language models (LLMs) by matching their unstructured outputs against standard answers, but also serves as the reward model to guide LLM optimization. Most evaluation frameworks rely on regularized matching or employ general LLMs for answer verification, which demands extensive, repetitive customization for regex rules or evaluation prompts. Two fundamental limitations persist in current methodologies: 1) the absence of comprehensive benchmarks that systematically evaluate verification capabilities across different LLMs; and 2) the nascent stage of verifier development, where existing approaches lack both the robustness to handle complex edge cases and the generalizability across different domains. In this work, we develop CompassVerifier, an accurate and robust lightweight verifier model for evaluation and outcome reward. It demonstrates multi-domain competency spanning math, knowledge, and diverse reasoning tasks, with the capability to process various answer types including multi-subproblems, formulas, and sequence answers, while effectively identifying abnormal/invalid responses. We introduce VerifierBench benchmark comprising model outputs collected from multiple data sources, augmented through manual analysis of meta error patterns to enhance CompassVerifier. We anticipate that CompassVerifier and VerifierBench will facilitate evaluation protocols and reinforcement learning research.

Learning logic rules for knowledge graph reasoning is essential as such rules provide interpretable explanations for reasoning and can be generalized to different domains. However, existing methods often face challenges such as searching in a vast search space (e.g., enumeration of relational paths or multiplication of high-dimensional matrices) and inefficient optimization (e.g., techniques based on reinforcement learning or EM algorithm). To address these limitations, this paper proposes a novel framework called LatentLogic to efficiently mine logic rules by controllable generation in the latent space. Specifically, to map the discrete relational paths into the latent space, we leverage a pre-trained VAE and employ a discriminator to establish an energy-based distribution. Additionally, we incorporate a sampler based on ordinary differential equations, enabling the efficient generation of logic rules in our approach. Extensive experiments on benchmark datasets demonstrate the effectiveness and efficiency of our proposed method.

Labeling large amounts of extractive summarization data is often prohibitive expensive due to time, financial, and expertise constraints, which poses great challenges to incorporating summarization system in practical applications. This limitation can be overcome by semi-supervised approaches: consistency-training and pseudo-labeling to make full use of unlabeled data. Researches on the two, however, are conducted independently, and very few works try to connect them. In this paper, we first use the noise-injected consistency training paradigm to regularize model predictions. Subsequently, we propose a novel entropy-constrained pseudo labeling strategy to obtain high-confidence labels from unlabeled predictions, which can obtain high-confidence labels from unlabeled predictions by comparing the entropy of supervised and unsupervised predictions. By combining consistency training and pseudo-labeling, this framework enforce a low-density separation between classes, which decently improves the performance of supervised learning over an insufficient labeled extractive summarization dataset.