CyberSec Research

We introduce our first reflective generative model MetaStone-S1, which obtains OpenAI o3's performance via the self-supervised process reward model (SPRM). Through sharing the backbone network and using task-specific heads for next token prediction and process scoring respectively, SPRM successfully integrates the policy model and process reward model(PRM) into a unified interface without extra process annotation, reducing over 99% PRM parameters for efficient reasoning. Equipped with SPRM, MetaStone-S1 is naturally suitable for test time scaling (TTS), and we provide three reasoning effort modes (low, medium, and high), based on the controllable thinking length. Moreover, we empirically establish a scaling law that reveals the relationship between total thinking computation and TTS performance. Experiments demonstrate that our MetaStone-S1 achieves comparable performance to OpenAI-o3-mini's series with only 32B parameter size. To support the research community, we have open-sourced MetaStone-S1 at https://github.com/MetaStone-AI/MetaStone-S1.

Large language models (LLMs) are excellent at maintaining high-level, convincing dialogues. They are being fast deployed as chatbots and evaluators in sensitive areas, such as peer review and mental health applications. This, along with the disparate accounts on their reasoning capabilities, calls for a closer examination of LLMs and their comprehension of dialogue. In this work we begin by evaluating LLMs' ability to maintain a debate--one of the purest yet most complex forms of human communication. Then we measure how this capability relates to their understanding of what is being talked about, namely, their comprehension of dialogical structures and the pragmatic context. We find that LLMs are capable of maintaining coherent, persuasive debates, often swaying the beliefs of participants and audiences alike. We also note that awareness or suspicion of AI involvement encourage people to be more critical of the arguments made. When polling LLMs on their comprehension of deeper structures of dialogue, however, they cannot demonstrate said understanding. Our findings tie the shortcomings of LLMs-as-evaluators to their (in)ability to understand the context. More broadly, for the field of argumentation theory we posit that, if an agent can convincingly maintain a dialogue, it is not necessary for it to know what it is talking about. Hence, the modelling of pragmatic context and coherence are secondary to effectiveness.

In recent years, neural models trained on large multilingual text and speech datasets have shown great potential for supporting low-resource languages. This study investigates the performances of two state-of-the-art Automatic Speech Recognition (ASR) models, OpenAI's Whisper (Small & Large-V2) and Facebook's Wav2Vec-BERT on Bangla, a low-resource language. We have conducted experiments using two publicly available datasets: Mozilla Common Voice-17 and OpenSLR to evaluate model performances. Through systematic fine-tuning and hyperparameter optimization, including learning rate, epochs, and model checkpoint selection, we have compared the models based on Word Error Rate (WER), Character Error Rate (CER), Training Time, and Computational Efficiency. The Wav2Vec-BERT model outperformed Whisper across all key evaluation metrics, demonstrated superior performance while requiring fewer computational resources, and offered valuable insights to develop robust speech recognition systems in low-resource linguistic settings.

Natural language generation (NLG) is increasingly deployed in high-stakes domains, yet common intrinsic evaluation methods, such as n-gram overlap or sentence plausibility, weakly correlate with actual decision-making efficacy. We propose a decision-oriented framework for evaluating generated text by directly measuring its influence on human and large language model (LLM) decision outcomes. Using market digest texts--including objective morning summaries and subjective closing-bell analyses--as test cases, we assess decision quality based on the financial performance of trades executed by human investors and autonomous LLM agents informed exclusively by these texts. Our findings reveal that neither humans nor LLM agents consistently surpass random performance when relying solely on summaries. However, richer analytical commentaries enable collaborative human-LLM teams to outperform individual human or agent baselines significantly. Our approach underscores the importance of evaluating generated text by its ability to facilitate synergistic decision-making between humans and LLMs, highlighting critical limitations of traditional intrinsic metrics.

Recent work has shown that distilling reasoning traces from a larger teacher model via supervised finetuning outperforms reinforcement learning with the smaller student model alone (Guo et al. 2025). However, there has not been a systematic study of what kind of reasoning demonstrations from the teacher are most effective in improving the student model's reasoning capabilities. In this work we curate high-quality "NaturalThoughts" by selecting reasoning traces from a strong teacher model based on a large pool of questions from NaturalReasoning (Yuan et al. 2025). We first conduct a systematic analysis of factors that affect distilling reasoning capabilities, in terms of sample efficiency and scalability for general reasoning tasks. We observe that simply scaling up data size with random sampling is a strong baseline with steady performance gains. Further, we find that selecting difficult examples that require more diverse reasoning strategies is more sample-efficient to transfer the teacher model's reasoning skills. Evaluated on both Llama and Qwen models, training with NaturalThoughts outperforms existing reasoning datasets such as OpenThoughts, LIMO, etc. on general STEM reasoning benchmarks including GPQA-Diamond, MMLU-Pro and SuperGPQA.

Reinforcement Learning from Human Feedback (RLHF) has emerged as a powerful technique for aligning large language models (LLMs) with human preferences. However, effectively aligning LLMs with diverse human preferences remains a significant challenge, particularly when they are conflict. To address this issue, we frame human value alignment as a multi-objective optimization problem, aiming to maximize a set of potentially conflicting objectives. We introduce Gradient-Adaptive Policy Optimization (GAPO), a novel fine-tuning paradigm that employs multiple-gradient descent to align LLMs with diverse preference distributions. GAPO adaptively rescales the gradients for each objective to determine an update direction that optimally balances the trade-offs between objectives. Additionally, we introduce P-GAPO, which incorporates user preferences across different objectives and achieves Pareto solutions that better align with the user's specific needs. Our theoretical analysis demonstrates that GAPO converges towards a Pareto optimal solution for multiple objectives. Empirical results on Mistral-7B show that GAPO outperforms current state-of-the-art methods, achieving superior performance in both helpfulness and harmlessness.

Recent advancements in artificial intelligence (AI), particularly in large language models (LLMs) such as OpenAI-o1 and DeepSeek-R1, have demonstrated remarkable capabilities in complex domains such as logical reasoning and experimental coding. Motivated by these advancements, numerous studies have explored the application of AI in the innovation process, particularly in the context of scientific research. These AI technologies primarily aim to develop systems that can autonomously conduct research processes across a wide range of scientific disciplines. Despite these significant strides, a comprehensive survey on AI for Research (AI4Research) remains absent, which hampers our understanding and impedes further development in this field. To address this gap, we present a comprehensive survey and offer a unified perspective on AI4Research. Specifically, the main contributions of our work are as follows: (1) Systematic taxonomy: We first introduce a systematic taxonomy to classify five mainstream tasks in AI4Research. (2) New frontiers: Then, we identify key research gaps and highlight promising future directions, focusing on the rigor and scalability of automated experiments, as well as the societal impact. (3) Abundant applications and resources: Finally, we compile a wealth of resources, including relevant multidisciplinary applications, data corpora, and tools. We hope our work will provide the research community with quick access to these resources and stimulate innovative breakthroughs in AI4Research.

Pruning is a highly effective approach for compressing large language models (LLMs), significantly reducing inference latency. However, conventional training-free structured pruning methods often employ a heuristic metric that indiscriminately removes some attention heads across all pruning layers, without considering their positions within the network architecture. In this work, we propose a novel pruning algorithm that strategically prunes attention heads in the model's higher layers. Since the removal of attention heads can alter the magnitude of token representations, we introduce an adaptive rescaling parameter that calibrates the representation scale post-pruning to counteract this effect. We conduct comprehensive experiments on a wide range of LLMs, including LLaMA3.1-8B, Mistral-7B-v0.3, Qwen2-7B, and Gemma2-9B. Our evaluation includes both generation and discriminative tasks across 27 datasets. The results consistently demonstrate that our method outperforms existing structured pruning methods. This improvement is particularly notable in generation tasks, where our approach significantly outperforms existing baselines.

Large language models (LLMs) excel at reasoning tasks requiring long thought sequences for planning, reflection, and refinement. However, their substantial model size and high computational demands are impractical for widespread deployment. Yet, small language models (SLMs) often struggle to learn long-form CoT reasoning due to their limited capacity, a phenomenon we refer to as the "SLMs Learnability Gap". To address this, we introduce \textbf{Mi}d-\textbf{Co}T \textbf{T}eacher \textbf{A}ssistant Distillation (MiCoTAl), a framework for improving long CoT distillation for SLMs. MiCoTA employs intermediate-sized models as teacher assistants and utilizes intermediate-length CoT sequences to bridge both the capacity and reasoning length gaps. Our experiments on downstream tasks demonstrate that although SLMs distilled from large teachers can perform poorly, by applying MiCoTA, they achieve significant improvements in reasoning performance. Specifically, Qwen2.5-7B-Instruct and Qwen2.5-3B-Instruct achieve an improvement of 3.47 and 3.93 respectively on average score on AIME2024, AMC, Olympiad, MATH-500 and GSM8K benchmarks. To better understand the mechanism behind MiCoTA, we perform a quantitative experiment demonstrating that our method produces data more closely aligned with base SLM distributions. Our insights pave the way for future research into long-CoT data distillation for SLMs.

Existing language learning tools, even those powered by Large Language Models (LLMs), often lack support for polyglot learners to build linguistic connections across vocabularies in multiple languages, provide limited customization for individual learning paces or needs, and suffer from detrimental cognitive offloading. To address these limitations, we design Do-It-Yourself Multilingual Knowledge Graph (DIY-MKG), an open-source system that supports polyglot language learning. DIY-MKG allows the user to build personalized vocabulary knowledge graphs, which are constructed by selective expansion with related words suggested by an LLM. The system further enhances learning through rich annotation capabilities and an adaptive review module that leverages LLMs for dynamic, personalized quiz generation. In addition, DIY-MKG allows users to flag incorrect quiz questions, simultaneously increasing user engagement and providing a feedback loop for prompt refinement. Our evaluation of LLM-based components in DIY-MKG shows that vocabulary expansion is reliable and fair across multiple languages, and that the generated quizzes are highly accurate, validating the robustness of DIY-MKG.

The rapid advancement of Large Language Models (LLMs) has intensified the need for evaluation frameworks that go beyond English centric benchmarks and address the requirements of linguistically diverse regions such as India. We present EKA-EVAL, a unified and production-ready evaluation framework that integrates over 35 benchmarks, including 10 Indic-specific datasets, spanning categories like reasoning, mathematics, tool use, long-context understanding, and reading comprehension. Compared to existing Indian language evaluation tools, EKA-EVAL offers broader benchmark coverage, with built-in support for distributed inference, quantization, and multi-GPU usage. Our systematic comparison positions EKA-EVAL as the first end-to-end, extensible evaluation suite tailored for both global and Indic LLMs, significantly lowering the barrier to multilingual benchmarking. The framework is open-source and publicly available at https://github.com/lingo-iitgn/ eka-eval and a part of ongoing EKA initiative (https://eka.soket.ai), which aims to scale up to over 100 benchmarks and establish a robust, multilingual evaluation ecosystem for LLMs.

As Large Language Models (LLMs) become increasingly widespread, understanding how specific training data shapes their outputs is crucial for transparency, accountability, privacy, and fairness. To explore how LLMs leverage and replicate their training data, we introduce a systematic approach centered on analyzing low-perplexity sequences - high-probability text spans generated by the model. Our pipeline reliably extracts such long sequences across diverse topics while avoiding degeneration, then traces them back to their sources in the training data. Surprisingly, we find that a substantial portion of these low-perplexity spans cannot be mapped to the corpus. For those that do match, we quantify the distribution of occurrences across source documents, highlighting the scope and nature of verbatim recall and paving a way toward better understanding of how LLMs training data impacts their behavior.

We present a comparative analysis of the parseability of structured outputs generated by small language models for open attribute-value extraction from clinical notes. We evaluate three widely used serialization formats: JSON, YAML, and XML, and find that JSON consistently yields the highest parseability. Structural robustness improves with targeted prompting and larger models, but declines for longer documents and certain note types. Our error analysis identifies recurring format-specific failure patterns. These findings offer practical guidance for selecting serialization formats and designing prompts when deploying language models in privacy-sensitive clinical settings.

Low-Rank Adapters (LoRAs) have transformed the fine-tuning of Large Language Models (LLMs) by enabling parameter-efficient updates. However, their widespread adoption remains limited by the reliance on GPU-based training. In this work, we propose a theoretically grounded approach to LoRA fine-tuning designed specifically for users with limited computational resources, particularly those restricted to standard laptop CPUs. Our method learns a meta-operator that maps any input dataset, represented as a probability distribution, to a set of LoRA weights by leveraging a large bank of pre-trained adapters for the Mistral-7B-Instruct-v0.2 model. Instead of performing new gradient-based updates, our pipeline constructs adapters via lightweight combinations of existing LoRAs directly on CPU. While the resulting adapters do not match the performance of GPU-trained counterparts, they consistently outperform the base Mistral model on downstream tasks, offering a practical and accessible alternative to traditional GPU-based fine-tuning.

Automatic medical coding has the potential to ease documentation and billing processes. For this task, transparency plays an important role for medical coders and regulatory bodies, which can be achieved using explainability methods. However, the evaluation of these approaches has been mostly limited to short text and binary settings due to a scarcity of annotated data. Recent efforts by Cheng et al. (2023) have introduced the MDACE dataset, which provides a valuable resource containing code evidence in clinical records. In this work, we conduct an in-depth analysis of the MDACE dataset and perform plausibility evaluation of current explainable medical coding systems from an applied perspective. With this, we contribute to a deeper understanding of automatic medical coding and evidence extraction. Our findings reveal that ground truth evidence aligns with code descriptions to a certain degree. An investigation into state-of-the-art approaches shows a high overlap with ground truth evidence. We propose match measures and highlight success and failure cases. Based on our findings, we provide recommendations for developing and evaluating explainable medical coding systems.

AI models are increasingly required to be multimodal, integrating disparate input streams into a coherent state representation on which subsequent behaviors and actions can be based. This paper seeks to understand how such models behave when input streams present conflicting information. Focusing specifically on vision-language models, we provide inconsistent inputs (e.g., an image of a dog paired with the caption "A photo of a cat") and ask the model to report the information present in one of the specific modalities (e.g., "What does the caption say / What is in the image?"). We find that models often favor one modality over the other, e.g., reporting the image regardless of what the caption says, but that different models differ in which modality they favor. We find evidence that the behaviorally preferred modality is evident in the internal representational structure of the model, and that specific attention heads can restructure the representations to favor one modality over the other. Moreover, we find modality-agnostic "router heads" which appear to promote answers about the modality requested in the instruction, and which can be manipulated or transferred in order to improve performance across datasets and modalities. Together, the work provides essential steps towards identifying and controlling if and how models detect and resolve conflicting signals within complex multimodal environments.

Language models can distinguish between testing and deployment phases -- a capability known as evaluation awareness. This has significant safety and policy implications, potentially undermining the reliability of evaluations that are central to AI governance frameworks and voluntary industry commitments. In this paper, we study evaluation awareness in Llama-3.3-70B-Instruct. We show that linear probes can separate real-world evaluation and deployment prompts, suggesting that current models internally represent this distinction. We also find that current safety evaluations are correctly classified by the probes, suggesting that they already appear artificial or inauthentic to models. Our findings underscore the importance of ensuring trustworthy evaluations and understanding deceptive capabilities. More broadly, our work showcases how model internals may be leveraged to support blackbox methods in safety audits, especially for future models more competent at evaluation awareness and deception.

Data quality is a critical driver of large language model performance, yet existing model-based selection methods focus almost exclusively on English. We introduce MuRating, a scalable framework that transfers high-quality English data-quality signals into a single rater for 17 target languages. MuRating aggregates multiple English "raters" via pairwise comparisons to learn unified document-quality scores,then projects these judgments through translation to train a multilingual evaluator on monolingual, cross-lingual, and parallel text pairs. Applied to web data, MuRating selects balanced subsets of English and multilingual content to pretrain a 1.2 B-parameter LLaMA model. Compared to strong baselines, including QuRater, AskLLM, DCLM and so on, our approach boosts average accuracy on both English benchmarks and multilingual evaluations, with especially large gains on knowledge-intensive tasks. We further analyze translation fidelity, selection biases, and underrepresentation of narrative material, outlining directions for future work.