ReSURE: Regularizing Supervision Unreliability for Multi-turn Dialogue Fine-tuning

Fine-tuning multi-turn dialogue systems requires high-quality supervision but often suffers from degraded performance when exposed to low-quality data. Supervision errors in early turns can propagate across subsequent turns, undermining coherence and response quality. Existing methods typically address data quality via static prefiltering, which decouples quality control from training and fails to mitigate turn-level error propagation. In this context, we propose ReSURE (Regularizing Supervision UnREliability), an adaptive learning method that dynamically down-weights unreliable supervision without explicit filtering. ReSURE estimates per-turn loss distributions using Welford's online statistics and reweights sample losses on the fly accordingly. Experiments on both single-source and mixed-quality datasets show improved stability and response quality. Notably, ReSURE enjoys positive Spearman correlations (0.21 ~ 1.0 across multiple benchmarks) between response scores and number of samples regardless of data quality, which potentially paves the way for utilizing large-scale data effectively. Code is publicly available at https://github.com/Elvin-Yiming-Du/ReSURE_Multi_Turn_Training.

Cultivating Helpful, Personalized, and Creative AI Tutors: A Framework for Pedagogical Alignment using Reinforcement Learning

The integration of large language models (LLMs) into education presents unprecedented opportunities for scalable personalized learning. However, standard LLMs often function as generic information providers, lacking alignment with fundamental pedagogical principles such as helpfulness, student-centered personalization, and creativity cultivation. To bridge this gap, we propose EduAlign, a novel framework designed to guide LLMs toward becoming more effective and responsible educational assistants. EduAlign consists of two main stages. In the first stage, we curate a dataset of 8k educational interactions and annotate them-both manually and automatically-along three key educational dimensions: Helpfulness, Personalization, and Creativity (HPC). These annotations are used to train HPC-RM, a multi-dimensional reward model capable of accurately scoring LLM outputs according to these educational principles. We further evaluate the consistency and reliability of this reward model. In the second stage, we leverage HPC-RM as a reward signal to fine-tune a pre-trained LLM using Group Relative Policy Optimization (GRPO) on a set of 2k diverse prompts. We then assess the pre- and post-finetuning models on both educational and general-domain benchmarks across the three HPC dimensions. Experimental results demonstrate that the fine-tuned model exhibits significantly improved alignment with pedagogical helpfulness, personalization, and creativity stimulation. This study presents a scalable and effective approach to aligning LLMs with nuanced and desirable educational traits, paving the way for the development of more engaging, pedagogically aligned AI tutors.

YH-MINER: Multimodal Intelligent System for Natural Ecological Reef Metric Extraction

Coral reefs, crucial for sustaining marine biodiversity and ecological processes (e.g., nutrient cycling, habitat provision), face escalating threats, underscoring the need for efficient monitoring. Coral reef ecological monitoring faces dual challenges of low efficiency in manual analysis and insufficient segmentation accuracy in complex underwater scenarios. This study develops the YH-MINER system, establishing an intelligent framework centered on the Multimodal Large Model (MLLM) for "object detection-semantic segmentation-prior input". The system uses the object detection module (mAP@0.5=0.78) to generate spatial prior boxes for coral instances, driving the segment module to complete pixel-level segmentation in low-light and densely occluded scenarios. The segmentation masks and finetuned classification instructions are fed into the Qwen2-VL-based multimodal model as prior inputs, achieving a genus-level classification accuracy of 88% and simultaneously extracting core ecological metrics. Meanwhile, the system retains the scalability of the multimodal model through standardized interfaces, laying a foundation for future integration into multimodal agent-based underwater robots and supporting the full-process automation of "image acquisition-prior generation-real-time analysis".

SynthDoc: Bilingual Documents Synthesis for Visual Document Understanding

This paper introduces SynthDoc, a novel synthetic document generation pipeline designed to enhance Visual Document Understanding (VDU) by generating high-quality, diverse datasets that include text, images, tables, and charts. Addressing the challenges of data acquisition and the limitations of existing datasets, SynthDoc leverages publicly available corpora and advanced rendering tools to create a comprehensive and versatile dataset. Our experiments, conducted using the Donut model, demonstrate that models trained with SynthDoc's data achieve superior performance in pre-training read tasks and maintain robustness in downstream tasks, despite language inconsistencies. The release of a benchmark dataset comprising 5,000 image-text pairs not only showcases the pipeline's capabilities but also provides a valuable resource for the VDU community to advance research and development in document image recognition. This work significantly contributes to the field by offering a scalable solution to data scarcity and by validating the efficacy of end-to-end models in parsing complex, real-world documents.

Offline RLHF Methods Need More Accurate Supervision Signals

With the rapid advances in Large Language Models (LLMs), aligning LLMs with human preferences become increasingly important. Although Reinforcement Learning with Human Feedback (RLHF) proves effective, it is complicated and highly resource-intensive. As such, offline RLHF has been introduced as an alternative solution, which directly optimizes LLMs with ranking losses on a fixed preference dataset. Current offline RLHF only captures the ``ordinal relationship'' between responses, overlooking the crucial aspect of ``how much'' one is preferred over the others. To address this issue, we propose a simple yet effective solution called \textbf{R}eward \textbf{D}ifference \textbf{O}ptimization, shorted as \textbf{RDO}. Specifically, we introduce {\it reward difference coefficients} to reweigh sample pairs in offline RLHF. We then develop a {\it difference model} involving rich interactions between a pair of responses for predicting these difference coefficients. Experiments with 7B LLMs on the HH and TL;DR datasets substantiate the effectiveness of our method in both automatic metrics and human evaluation, thereby highlighting its potential for aligning LLMs with human intent and values.

CMR Scaling Law: Predicting Critical Mixture Ratios for Continual Pre-training of Language Models

Large Language Models (LLMs) excel in diverse tasks but often underperform in specialized fields due to limited domain-specific or proprietary corpus. Continual pre-training (CPT) enhances LLM capabilities by imbuing new domain-specific or proprietary knowledge while replaying general corpus to prevent catastrophic forgetting. The data mixture ratio of general corpus and domain-specific corpus, however, has been chosen heuristically, leading to sub-optimal training efficiency in practice. In this context, we attempt to re-visit the scaling behavior of LLMs under the hood of CPT, and discover a power-law relationship between loss, mixture ratio, and training tokens scale. We formalize the trade-off between general and domain-specific capabilities, leading to a well-defined Critical Mixture Ratio (CMR) of general and domain data. By striking the balance, CMR maintains the model's general ability and achieves the desired domain transfer, ensuring the highest utilization of available resources. Therefore, if we value the balance between efficiency and effectiveness, CMR can be consider as the optimal mixture ratio.Through extensive experiments, we ascertain the predictability of CMR, and propose CMR scaling law and have substantiated its generalization. These findings offer practical guidelines for optimizing LLM training in specialized domains, ensuring both general and domain-specific performance while efficiently managing training resources.

SimCT: A Simple Consistency Test Protocol in LLMs Development Lifecycle

In this work, we report our efforts to advance the standard operation procedure of developing Large Language Models (LLMs) or LLMs-based systems or services in industry. We introduce the concept of Large Language Model Development Lifecycle (LDLC) and then highlight the importance of consistency test in ensuring the delivery quality. The principled solution of consistency test, however, is usually overlooked by industrial practitioners and not urgent in academia, and current practical solutions are insufficiently rigours and labor-intensive. We thus propose a simple yet effective consistency test protocol, named SimCT. SimCT is mainly to proactively check the consistency across different development stages of "bare metal" LLMs or associated services without accessing the model artifacts, in an attempt to expedite the delivery by reducing the back-and-forth alignment communications among multiple teams involved in different development stages. Specifically, SimCT encompasses response-wise and model-wise tests. We implement the protocol with LightGBM and Student's t-test for two components respectively, and perform extensive experiments to substantiate the effectiveness of SimCT and the involved components.

What Makes Good Few-shot Examples for Vision-Language Models?

Despite the notable advancements achieved by leveraging pre-trained vision-language (VL) models through few-shot tuning for downstream tasks, our detailed empirical study highlights a significant dependence of few-shot learning outcomes on the careful selection of training examples - a facet that has been previously overlooked in research. In this study, we delve into devising more effective strategies for the meticulous selection of few-shot training examples, as opposed to relying on random sampling, to enhance the potential of existing few-shot prompt learning methodologies. To achieve this, we assess the effectiveness of various Active Learning (AL) techniques for instance selection, such as Entropy and Margin of Confidence, within the context of few-shot training. Furthermore, we introduce two innovative selection methods - Representativeness (REPRE) and Gaussian Monte Carlo (Montecarlo) - designed to proactively pinpoint informative examples for labeling in relation to pre-trained VL models. Our findings demonstrate that both REPRE and Montecarlo significantly surpass both random selection and AL-based strategies in few-shot training scenarios. The research also underscores that these instance selection methods are model-agnostic, offering a versatile enhancement to a wide array of few-shot training methodologies.

Balancing Speciality and Versatility: a Coarse to Fine Framework for Supervised Fine-tuning Large Language Model

Aligned Large Language Models (LLMs) showcase remarkable versatility, capable of handling diverse real-world tasks. Meanwhile, aligned LLMs are also expected to exhibit speciality, excelling in specific applications. However, fine-tuning with extra data, a common practice to gain speciality, often leads to catastrophic forgetting (CF) of previously acquired versatility, hindering the model's performance across diverse tasks. In response to this challenge, we propose CoFiTune, a coarse to fine framework in an attempt to strike the balance between speciality and versatility. At the coarse-grained level, an empirical tree-search algorithm is utilized to pinpoint and update specific modules that are crucial for speciality, while keeping other parameters frozen; at the fine-grained level, a soft-masking mechanism regulates the update to the LLMs, mitigating the CF issue without harming speciality. In an overall evaluation of both speciality and versatility, CoFiTune consistently outperforms baseline methods across diverse tasks and model scales. Compared to the full-parameter SFT, CoFiTune leads to about 14% versatility improvement and marginal speciality loss on a 13B model. Lastly, based on further analysis, we provide a speculative insight into the information forwarding process in LLMs, which helps explain the effectiveness of the proposed method. The code is available at https://github.com/rattlesnakey/CoFiTune.

Consistency Matters: Explore LLMs Consistency From a Black-Box Perspective

Nowadays both commercial and open-source academic LLM have become the mainstream models of NLP. However, there is still a lack of research on LLM consistency, meaning that throughout the various stages of LLM research and deployment, its internal parameters and capabilities should remain unchanged. This issue exists in both the industrial and academic sectors. The solution to this problem is often time-consuming and labor-intensive, and there is also an additional cost of secondary deployment, resulting in economic and time losses. To fill this gap, we build an LLM consistency task dataset and design several baselines. Additionally, we choose models of diverse scales for the main experiments. Specifically, in the LightGBM experiment, we used traditional NLG metrics (i.e., ROUGE, BLEU, METEOR) as the features needed for model training. The final result exceeds the manual evaluation and GPT3.5 as well as other models in the main experiment, achieving the best performance. In the end, we use the best performing LightGBM model as the base model to build the evaluation tool, which can effectively assist in the deployment of business models. Our code and tool demo are available at https://github.com/heavenhellchen/Consistency.git