Training on model-generated synthetic data is a promising approach for finetuning LLMs, but it remains unclear when it helps or hurts. In this paper, we investigate this question for math reasoning via an empirical study, followed by building a conceptual understanding of our observations. First, we find that while the typical approach of finetuning a model on synthetic correct or positive problem-solution pairs generated by capable models offers modest performance gains, sampling more correct solutions from the finetuned learner itself followed by subsequent fine-tuning on this self-generated data doubles the efficiency of the same synthetic problems. At the same time, training on model-generated positives can amplify various spurious correlations, resulting in flat or even inverse scaling trends as the amount of data increases. Surprisingly, we find that several of these issues can be addressed if we also utilize negative responses, i.e., model-generated responses that are deemed incorrect by a final answer verifier. Crucially, these negatives must be constructed such that the training can appropriately recover the utility or advantage of each intermediate step in the negative response. With this per-step scheme, we are able to attain consistent gains over only positive data, attaining performance similar to amplifying the amount of synthetic data by $\mathbf{8 \times}$. We show that training on per-step negatives can help to unlearn spurious correlations in the positive data, and is equivalent to advantage-weighted reinforcement learning (RL), implying that it inherits robustness benefits of RL over imitating positive data alone.

Abstract--Negative bias refers to the tendency of large language models (LLMs) to excessively generate negative responses in binary decision tasks (e.g., yes-no question answering). Previous research has focused on detecting and addressing negative attention heads that induce negative bias. However, the underlying detailed factors influencing negative bias remain underexplored. In this paper, we demonstrate that LLMs exhibit format-level negative bias, meaning the prompt format more influences their responses than the semantics of the negative response. For the fine-grained study of the negative bias, we introduce a pipeline for constructing the evaluation set, which systematically categorizes the dataset into three subsets based on the model's parametric knowledge: correct, incorrect, and insufficient relevant knowledge. Through analysis of this evaluation set, we identify a shortcut behavior in which models tend to generate negative responses when they lack sufficient knowledge to answer a yes-no question, leading to negative bias. We further examine how negative bias changes under various prompting scenarios related to parametric knowledge. We observe that providing relevant context and offering an "I don't know" option generally reduces negative bias, whereas chain-of-thought prompting tends to amplify the bias. Finally, we demonstrate that the degree of negative bias can vary depending on the type of prompt, which influences the direction of the response. Our work reveals the various factors that influence negative bias, providing critical insights for mitigating it in LLMs. ECENT advances in the capabilities and emergent abilities of large language models (LLMs) have led to rapid improvements in the performance of a wide range of natural language processing (NLP) tasks [1]-[5]. Leveraging their ability to follow instructions, LLMs are able to perform complex, previously unseen tasks, enabling human-like interactions [6]-[9]. One critical issue is the hallucination problem, where the model generates content that contains misleading information, which does not correspond to the given context or real-world knowledge [11]. J. Song was with the Department of Electrical and Computer Engineering at Seoul National University, South Korea (coms1580@gmail.com).

Large Language Models like GPT-4 adjust their responses not only based on the question asked, but also on how it is emotionally phrased. We systematically vary the emotional tone of 156 prompts - spanning controversial and everyday topics - and analyze how it affects model responses. Our findings show that GPT-4 is three times less likely to respond negatively to a negatively framed question than to a neutral one. This suggests a "rebound" bias where the model overcorrects, often shifting toward neutrality or positivity. On sensitive topics (e.g., justice or politics), this effect is even more pronounced: tone-based variation is suppressed, suggesting an alignment override. We introduce concepts like the "tone floor" - a lower bound in response negativity - and use tone-valence transition matrices to quantify behavior. Visualizations based on 1536-dimensional embeddings confirm semantic drift based on tone. Our work highlights an underexplored class of biases driven by emotional framing in prompts, with implications for AI alignment and trust. Code and data are available at: https://github.com/bardolfranck/llm-responses-viewer

Reinforcement Fine-Tuning (RFT) with verifiable rewards has advanced large language models but remains underexplored for Vision-Language (VL) models. The Vision-Language Reward Model (VL-RM) is key to aligning VL models by providing structured feedback, yet training effective VL-RMs faces two major challenges. First, the bootstrapping dilemma arises as high-quality training data depends on already strong VL models, creating a cycle where self-generated supervision reinforces existing biases. Second, modality bias and negative example amplification occur when VL models hallucinate incorrect visual attributes, leading to flawed preference data that further misguides training. To address these issues, we propose an iterative training framework leveraging vision experts, Chain-of-Thought (CoT) rationales, and Margin-based Rejection Sampling. Our approach refines preference datasets, enhances structured critiques, and iteratively improves reasoning. Experiments across VL-RM benchmarks demonstrate superior performance in hallucination detection and multimodal reasoning, advancing VL model alignment with reinforcement learning.

Large Language Models (LLMs) excel at many tasks but struggle with ambiguous scenarios where multiple valid responses exist, often yielding unreliable results. Conversely, Small Language Models (SLMs) demonstrate robustness in such scenarios but are susceptible to misleading or adversarial inputs. We observed that LLMs handle negative examples effectively, while SLMs excel with positive examples. To leverage their complementary strengths, we introduce SLIDE (Small and Large Integrated for Dialogue Evaluation), a method integrating SLMs and LLMs via adaptive weighting. Building on SLIDE, we further propose a Dual-Refinement Evaluation (DRE) method to enhance SLM-LLM integration: (1) SLM-generated insights guide the LLM to produce initial evaluations; (2) SLM-derived adjustments refine the LLM's scores for improved accuracy. Experiments demonstrate that DRE outperforms existing methods, showing stronger alignment with human judgment across diverse benchmarks. This work illustrates how combining small and large models can yield more reliable evaluation tools, particularly for open-ended tasks such as dialogue evaluation.

Training on model-generated synthetic data is a promising approach for finetuning LLMs, but it remains unclear when it helps or hurts. In this paper, we investigate this question for math reasoning via an empirical study, followed by building a conceptual understanding of our observations. First, we find that while the typical approach of finetuning a model on synthetic correct or positive problem-solution pairs generated by capable models offers modest performance gains, sampling more correct solutions from the finetuned learner itself followed by subsequent fine-tuning on this self-generated data doubles the efficiency of the same synthetic problems. At the same time, training on model-generated positives can amplify various spurious correlations, resulting in flat or even inverse scaling trends as the amount of data increases. Surprisingly, we find that several of these issues can be addressed if we also utilize negative responses, i.e., model-generated responses that are deemed incorrect by a final answer verifier. Crucially, these negatives must be constructed such that the training can appropriately recover the utility or advantage of each intermediate step in the negative response.

We introduce REFA, a family of reference-free alignment methods that optimize over multiple user preferences while enforcing fine-grained length control. Our approach integrates deviation-based weighting to emphasize high-quality responses more strongly, length normalization to prevent trivial short-response solutions, and an EOS-probability regularizer to mitigate dataset-induced brevity biases. Theoretically, we show that under the Uncertainty Reduction with Sequence Length Assertion (URSLA), naive length normalization can still incentivize length-based shortcuts. By contrast, REFA corrects these subtle incentives, guiding models toward genuinely more informative and higher-quality outputs. Empirically, REFA sets a new state-of-the-art among reference-free alignment methods, producing richer responses aligned more closely with human preferences. Compared to a base supervised fine-tuned (SFT) mistral-7b model that achieves 8.4% length-controlled win rate (LC-WR) and 6.2% win rate (WR), our best REFA configuration attains 21.62% LC-WR and 19.87% WR on the AlpacaEval v2 benchmark. This represents a substantial improvement over both the strongest multi-preference baseline, InfoNCA (16.82% LC-WR, 10.44% WR), and the strongest reference-free baseline, SimPO (20.01% LC-WR, 17.65% WR)

We introduce Simultaneous Weighted Preference Optimization (SWEPO), a novel extension of Direct Preference Optimization (DPO) designed to accommodate multiple dynamically chosen positive and negative responses for each query. SWEPO employs a weighted group contrastive loss, assigning weights to responses based on their deviation from the mean reward score. This approach effectively prioritizes responses that are significantly better or worse than the average, enhancing optimization. Our theoretical analysis demonstrates that simultaneously considering multiple preferences reduces alignment bias, resulting in more robust alignment. Additionally, we provide insights into the training dynamics of our loss function and a related function, InfoNCA. Empirical validation on the UltraFeedback dataset establishes SWEPO as state-of-the-art, with superior performance in downstream evaluations using the AlpacaEval dataset.

Recent research has demonstrated that Large Language Models (LLMs) are not limited to text-only tasks but can also function as multimodal models across various modalities, including audio, images, and videos. In particular, research on 3D Large Multimodal Models (3D LMMs) is making notable strides, driven by the potential of processing higher-dimensional data like point clouds. However, upon closer examination, we find that the visual and textual content within each sample of existing training datasets lacks both high informational granularity and clarity, which serve as a bottleneck for precise cross-modal understanding. To address these issues, we propose CL3DOR, Contrastive Learning for 3D large multimodal models via Odds ratio on high-Resolution point clouds, designed to ensure greater specificity and clarity in both visual and textual content. Specifically, we increase the density of point clouds per object and construct informative hard negative responses in the training dataset to penalize unwanted responses. To leverage hard negative responses, we incorporate the odds ratio as an auxiliary term for contrastive learning into the conventional language modeling loss. CL3DOR achieves state-of-the-art performance in 3D scene understanding and reasoning benchmarks. Additionally, we demonstrate the effectiveness of CL3DOR's key components through extensive experiments.

Instruction-following is a fundamental capability of language models, requiring the model to recognize even the most subtle requirements in the instructions and accurately reflect them in its output. Such an ability is well-suited for and often optimized by preference learning. Such practice can introduce content variations irrelevant to whether the instruction is precisely followed (e.g., different expressions about the same semantic), interfering with the goal of teaching models to recognize the key differences that lead to improved instruction following. R, a self-play framework integrating tree-search self-refinement to yield valid and comparable preference pairs free from distractions. By playing against itself, an LLM employs a tree-search strategy to refine its previous responses with respect to the instruction while minimizing unnecessary variations. R demonstrates promising scalability and transferability, greatly enhancing models like GLM-4-9B and LLaMA3-70B. We also identify how inference scaling in tree search would impact model performance. Our code and data are publicly available at https://github.com/thu-coai/SPaR. Write a story and end it with "The devil is in the details." The is in the details. Figure 1: An example of the interfering factors (story content) in independently sampled multiple responses (Left). Refined response pairs exclude these factors, highlight the key difference (ending sentence), and lead to improved performance on iteratively trained LLaMA3-8B-Instruct (Right).