Recent studies show that large language models (LLMs) exhibit self-preference bias when serving as judges, meaning they tend to favor their own responses over those generated by other models. Existing methods typically measure this bias by calculating the difference between the scores a judge model assigns to its own responses and those it assigns to responses from other models. However, this approach conflates self-preference bias with response quality, as higher-quality responses from the judge model may also lead to positive score differences, even in the absence of bias. To address this issue, we introduce gold judgments as proxies for the actual quality of responses and propose the DBG score, which measures self-preference bias as the difference between the scores assigned by the judge model to its own responses and the corresponding gold judgments. Since gold judgments reflect true response quality, the DBG score mitigates the confounding effect of response quality on bias measurement. Using the DBG score, we conduct comprehensive experiments to assess self-preference bias across LLMs of varying versions, sizes, and reasoning abilities. Additionally, we investigate two factors that influence and help alleviate self-preference bias: response text style and the post-training data of judge models. Finally, we explore potential underlying mechanisms of self-preference bias from an attention-based perspective. Our code and data are available at https://github.com/zhiyuanc2001/self-preference.

In-context learning (ICL) performance heavily relies on the quality and ordering of demonstrations. Iterative selection (IS) is a promising approach to address this issue, but existing IS methods face two key challenges: the oversimplification of process reward signals that guide intermediate steps (often using single-dimensional metrics) and the lack of outcome reward signals that directly optimize final-task accuracy (relying solely on binary terminal feedback like correct/incorrect predictions). To address these issues, we propose a reinforcement learning method R-Mix which models iterative demonstration selection as a Markov Decision Process (MDP), crafting hybrid reward signals — combining outcome-based accuracy signals (i.e., outcome rewards) with process-oriented signals (i.e, process rewards) like stepwise influence and label entropy improvement. Our analysis reveals a positive but trade-off relationship between outcome rewards and process rewards, underscoring the importance of both components for effective policy optimization. We further introduce a dual-head policy architecture that explicitly decouples input-semantic relevance and label-content compatibility. Experiments across NLP benchmarks demonstrate superior performance over state-of-the-art methods, with ablation studies validating the necessity of both reward components and architectural disentanglement. Our work has deeply explored the effective potential of ICL through demonstration selection.

In-Context Learning (ICL) is an essential emergent ability of Large Language Models (LLMs), and recent studies introduce CoT to exemplars of ICL to enhance the reasoning capability, especially in mathematics tasks. However, given the continuous advancement of model capabilities, it remains unclear whether CoT exemplars still benefit recent, stronger models in such tasks. Through systematic experiments, we find that for recent strong models such as the Qwen2.5 series, adding traditional CoT exemplars does not improve reasoning performance compared to Zero-Shot CoT. Instead, their primary function is to align the output format with human expectations. We further investigate the effectiveness of enhanced CoT exemplars, constructed using answers from advanced models such as Qwen2.5-Max and DeepSeek-R1. Experimental results indicate that these enhanced exemplars still fail to improve the model’s reasoning performance. Further analysis reveals that models tend to ignore the exemplars and focus primarily on the instructions, leading to no observable gain in reasoning ability. Overall, our findings highlight the limitations of the current ICL+CoT framework in mathematical reasoning, calling for a re-examination of the ICL paradigm and the definition of exemplars.

How can we construct an automated debate judge to evaluate an extensive, vibrant, multi-turn debate? This task is challenging, as judging a debate involves grappling with lengthy texts, intricate argument relationships, and multi-dimensional assessments.At the same time, current research mainly focuses on short dialogues, rarely touching upon the evaluation of an entire debate.In this paper, by leveraging Large Language Models (LLMs), we propose Debatrix, which makes the analysis and assessment of multi-turn debates more aligned with majority preferences. Specifically, Debatrix features a vertical, iterative chronological analysis and a horizontal, multi-dimensional evaluation collaboration.To align with real-world debate scenarios, we introduced the PanelBench benchmark, comparing our system’s performance to actual debate outcomes.The findings indicate a notable enhancement over directly using LLMs for debate evaluation.Source code and benchmark data are available at https://github.com/ljcleo/debatrix.

Syllogistic reasoning, a typical form of deductive reasoning, is a critical capability widely required in natural language understanding tasks, such as text entailment and question answering. To better facilitate research on syllogistic reasoning, we develop a benchmark called SylloBase that differs from existing syllogistic datasets in three aspects: (1) Covering a complete taxonomy of syllogism reasoning patterns; (2) Containing both automatically and manually constructed samples; and (3) Involving both the generation and understanding tasks. We automatically construct 50k template-based syllogism samples by mining syllogism patterns from Wikidata and ConceptNet. To improve our dataset’s naturalness and challenge, we apply GPT-3 to paraphrase the template-based data and further manually rewrite 1,000 samples as the test set. State-of-the-art pre-trained language models can achieve the best generation ROUGE-L of 38.72 by T5 and the best multi-choice accuracy of 72.77% by RoBERTa on SylloBase, which indicates the great challenge of learning diverse syllogistic reasoning types on SylloBase. Our datasets are released at https://github.com/casually-PYlearner/SYLLOBASE.