Reproducing research results from papers and released code is central to scientific progress. Existing works have introduced benchmarks to evaluate whether LLM agents can assist with reproducibility, but they are difficult to scale due to their reliance on substantial manual effort for data curation and evaluation. We introduce ReproRepo, a scalable framework for reproducibility evaluation that leverages human-raised GitHub issues as naturally occurring supervision on realistic reproduction blockers. We instantiate ReproRepo on 1,149 recent machine learning papers from major conferences and evaluate four frontier model-agent configurations. Our results show that LLM agents, even without executing code, can identify many real-world reproducibility problems from paper-repository pairs: the best agent in our study, namely Codex with GPT-5.5, surfaces at least one semantically related human-reported blocker for ∼90% of papers in the study. Further analysis shows that agents are particularly effective for surfacing visible failures and identifying the right semantic region, but may still be insufficient in exact localization. ReproRepo can serve as a reusable, scalable framework for future evaluations of LLM agents on real-world reproducibility auditing. Our code is released at https://github.com/LithiumDA/ReproRepo.

As the diversity of users increases, the capability of providing personalized responses by large language models (LLMs) has become increasingly important. Existing approaches have only limited successes in LLM personalization, due to the absence of personalized learning or the reliance on shared personal data. This paper proposes a new approach for a few-shot personalization of LLMs with their mis-aligned responses (Fermi). Our key idea is to learn a set of personalized prompts for each user by progressively improving the prompts using LLMs, based on user profile (e.g., demographic information) and a few examples of previous opinions. During an iterative process of prompt improvement, we incorporate the contexts of mis-aligned responses by LLMs, which are especially crucial for the effective personalization of LLMs. In addition, we develop an effective inference method to further leverage the context of the test query and the personalized prompts. Our experimental results demonstrate that Fermi significantly improves performance across various benchmarks, compared to best-performing baselines.

Preference modeling techniques, such as direct preference optimization (DPO), has shown effective in enhancing the generalization abilities of large language model (LLM). However, in tasks involving video instruction-following, providing informative feedback, especially for open-ended conversations, remains a significant challenge. While previous studies have explored using large multimodal models (LMMs) as reward models for guiding preference modeling, their ability to accurately assess the quality of generated responses and their alignment with video content has not been conclusively demonstrated. This paper introduces a novel framework that utilizes detailed video captions as a proxy of video content, enabling language models to incorporate this information as supporting evidence for scoring video Question Answering (QA) predictions. Our approach demonstrates robust alignment with OpenAI GPT-4V model’s reward mechanism, which directly takes video frames as input. Furthermore, we show that applying our reward mechanism to DPO algorithm significantly improves model performance on open-ended video QA tasks.

Current long chain-of-thought (long-CoT) models excel at mathematical reasoning but rely on slow and error-prone natural language traces. Tool-augmented agents address arithmetic via code execution, but often falter on complex logical tasks. We introduce a fine-tuning framework, DualDistill, that distills complementary reasoning strategies from multiple teachers into a unified student model. Using this approach, we train Agentic-R1, which dynamically selects the optimal strategy for each query, invoking tools for arithmetic and algorithmic problems and using text-based reasoning for abstract ones. Our method improves accuracy on computation-intensive tasks and reduces inference latency on standard benchmarks, demonstrating the promise of multi-strategy distillation for robust and efficient reasoning.

Chain-of-thought (CoT) reasoning in vision language models (VLMs) is crucial for improving interpretability and trustworthiness. However, current training recipes often relying on datasets dominated by short annotations with minimal rationales. In this work, we show that training VLM on short answers leads to poor generalization on reasoning tasks that require more detailed explanations. To address this limitation, we propose a two-stage post-training strategy that extends the usage of short answer data for enhanced CoT reasoning. First, we augment short answers with CoT reasoning generated by GPT-4o, enhancing the VLM’s CoT capabilities through fine-tuning. Second, we leverage short answers as outcome rewards for reinforcement learning. Specifically, short answers are used as correctness indicators to construct positive (correct) and negative (incorrect) pairs from model-generated reasoning chains. These pairs are then used to calibrate the model’s reasoning via Direct Preference Optimization. Our experiments show significant improvements in CoT reasoning on benchmark datasets, along with enhanced generalization to direct answer prediction. This work provides a critical data resource for VLM CoT training and demonstrates the effectiveness of outcome rewards for multimodal models post-training.

Large Multimodal Models (LMM) are built across modalities and the misalignment between two modalities can result in “hallucination”, generating textual outputs that are not grounded by the multimodal information in context. To address the multimodal misalignment issue, we adapt the Reinforcement Learning from Human Feedback (RLHF) from the text domain to the vision-language alignment, where human annotators are asked to compare two responses and pinpoint the more hallucinated one, and the vision-language model is trained to maximize the simulated human rewards. We propose a new alignment algorithm called Factually Augmented RLHF that augments the reward model with additional factual information such as image captions and ground-truth multi-choice options, which alleviates the reward hacking phenomenon in RLHF and further improves the performance. We also enhance the GPT-4-generated training data (for vision instruction tuning) with previously available human-written image-text pairs to improve the general capabilities of our model. To evaluate the proposed approach in real-world scenarios, we develop a new evaluation benchmark MMHAL-BENCH with a special focus on penalizing hallucinations. As the first LMM trained with RLHF, our approach achieves remarkable improvement on the LLaVA-Bench dataset with the 96% performance level of the text-only GPT-4 (while previous best methods can only achieve the 87% level), and an improvement of 60% on MMHAL-BENCH over other baselines.

An open challenge in recent machine learning is about how to improve the reasoning capability of large language models (LLMs) in a black-box setting, i.e., without access to detailed information such as output token probabilities. Existing approaches either rely on accessibility (which is often unrealistic) or involve significantly increased train- and inference-time costs. This paper addresses those limitations or shortcomings by proposing a novel approach, namely CoBB (Correct for improving QA reasoning of Black-Box LLMs). It uses a trained adaptation model to perform a seq2seq mapping from the often-imperfect reasonings of the original black-box LLM to the correct or improved reasonings. Specifically, the adaptation model is initialized with a relatively small open-source LLM and adapted over a collection of sub-sampled training pairs. To select the representative pairs of correct and incorrect reasonings, we formulated the dataset construction as an optimization problem that minimizes the statistical divergence between the sampled subset and the entire collection, and solved it via a genetic algorithm. We then train the adaptation model over the sampled pairs by contrasting the likelihoods of correct and incorrect reasonings. Our experimental results demonstrate that CoBB significantly improves reasoning accuracy across various QA benchmarks, compared to the best-performing adaptation baselines.

The remarkable performance of large language models (LLMs) in zero-shot language understanding has garnered significant attention.However, employing LLMs for large-scale inference or domain-specific fine-tuning requires immense computational resources due to their substantial model size. To overcome these limitations, we introduce a novel method, namely GenCo, which leverages the strong generative power of LLMs to assist in training a smaller and more adaptable language model. In our method, an LLM plays an important role in the self-training loop of a smaller model in two important ways. Firstly, we utilize an LLM to generate multiple augmented texts for each input instance to enhance its semantic meaning for better understanding. Secondly, we additionally generate high-quality training instances conditioned on predicted labels, ensuring the generated texts are relevant to the labels. In this way, GenCo not only corrects the errors of predicted labels during self-training but also eliminates the need for extensive unlabeled texts. In our experiments, GenCo outperforms previous state-of-the-art methods when only limited (<5% of original) in-domain text data is available. Notably, our approach surpasses Alpaca-7B with human instructions, highlighting the significance of self-training.

Large language models (LMs), while powerful, are not immune to mistakes, but can be difficult to retrain. Our goal is for an LM to continue to improve after deployment, without retraining, using feedback from the user. Our approach pairs an LM with (i) a growing memory of cases where the user identified an output error and provided general feedback on how to correct it (ii) a corrector model, trained to translate this general feedback into specific edits to repair the model output. Given a new, unseen input, our model can then use feedback from similar, past cases to repair output errors that may occur. We instantiate our approach using an existing, fixed model for script generation, that takes a goal (e.g., “bake a cake”) and generates a partially ordered sequence of actions to achieve that goal, sometimes containing errors. Our memory-enhanced system, , learns to apply user feedback to repair such errors (up to 30 points improvement), while making a start at avoiding similar past mistakes on new, unseen examples (up to 7 points improvement in a controlled setting). This is a first step towards strengthening deployed models, potentially broadening their utility. Our code and data is available at https://github.com/allenai/interscript

Predicting the effects of unexpected situations is an important reasoning task, e.g., would cloudy skies help or hinder plant growth? Given a context, the goal of such situational reasoning is to elicit the consequences of a new situation (st) that arises in that context. We propose CURIE, a method to iteratively build a graph of relevant consequences explicitly in a structured situational graph (st graph) using natural language queries over a finetuned language model. Across multiple domains, CURIE generates st graphs that humans find relevant and meaningful in eliciting the consequences of a new situation (75% of the graphs were judged correct by humans). We present a case study of a situation reasoning end task (WIQA-QA), where simply augmenting their input with st graphs improves accuracy by 3 points. We show that these improvements mainly come from a hard subset of the data, that requires background knowledge and multi-hop reasoning.

Current Open-Domain Question Answering (ODQA) models typically include a retrieving module and a reading module, where the retriever selects potentially relevant passages from open-source documents for a given question, and the reader produces an answer based on the retrieved passages. The recently proposed Fusion-in-Decoder (FiD) framework is a representative example, which is built on top of a dense passage retriever and a generative reader, achieving the state-of-the-art performance. In this paper we further improve the FiD approach by introducing a knowledge-enhanced version, namely KG-FiD. Our new model uses a knowledge graph to establish the structural relationship among the retrieved passages, and a graph neural network (GNN) to re-rank the passages and select only a top few for further processing. Our experiments on common ODQA benchmark datasets (Natural Questions and TriviaQA) demonstrate that KG-FiD can achieve comparable or better performance in answer prediction than FiD, with less than 40% of the computation cost.