Large reasoning models (LRMs) generate intermediate reasoning traces before producing final answers, yielding strong gains on multi-step and mathematical tasks. Yet aligning LRMs with human preferences, a crucial prerequisite for model deployment, remains underexplored. The statistically correct objective for preference alignment requires marginalizing over reasoning traces, but this computation is intractable in practice. A common workaround optimizes a single sampled trajectory, which introduces substantial gradient variance from stochastic trace sampling. To address this challenge, we frame preference optimization for LRMs through the lens of the bias--variance trade-off and propose Bias--Variance Optimized Preference Optimization (BVPO), a simple, drop-in method that mixes two gradient estimators: a high-variance trace-based estimator and a low-variance empty-trace estimator obtained by disabling reasoning trace generation. Our theory shows that BVPO strictly reduces trace-induced variance for any nontrivial mixture, provides a closed-form choice of the mixing weight that minimizes mean-squared error relative to the true marginal gradient, and under standard smoothness and step-size conditions, tightens classical convergence bounds for stochastic gradient descent. Empirically, BVPO improves alignment over the best baseline by up to 7.8 points on AlpacaEval~2 and 6.8 points on Arena-Hard. Despite being trained only on general conversational data, BVPO also boosts reasoning performance for base models by up to 4.0 points on the average of six math reasoning benchmarks. These results identify variance from trace sampling as a key bottleneck and demonstrate that directly optimizing the bias--variance trade-off yields more stable training and stronger overall performance.

Diffusion large language models (dLLMs) have recently emerged as a promising alternative to autoregressive (AR) models, offering advantages such as accelerated parallel decoding and bidirectional context modeling. However, the vanilla decoding strategy in discrete dLLMs suffers from a critical limitation: once a token is accepted, it can no longer be revised in subsequent steps. As a result, early mistakes persist across iterations, harming both intermediate predictions and final output quality. To address this issue, we propose Tolerator (Token-Level Cross-Validation Refinement), a training-free decoding strategy that leverages cross-validation among predicted tokens. Unlike existing methods that follow a single progressive unmasking procedure, Tolerator introduces a two-stage process: (i) sequence fill-up and (ii) iterative refinement by remasking and decoding a subset of tokens while treating the remaining as context. This design enables previously accepted tokens to be reconsidered and corrected when necessary, leading to more reliable diffusion decoding outputs. We evaluate Tolerator on five standard benchmarks covering language understanding, code generation, and mathematics. Experiments show that our method achieves consistent improvements over the baselines under the same computational budget. These findings suggest that decoding algorithms are crucial to realizing the full potential of diffusion large language models. Code and data are publicly available.

The promise of generative AI to revolutionize education is constrained by the pedagogical limits of large language models (LLMs). A major issue is the lack of access to high-quality training data that reflect the learning of actual students. Prompt engineering has emerged as a stopgap, but the ability of prompts to encode complex pedagogical strategies in rule-based natural language is inherently limited. To address this gap we introduce TeachLM - an LLM optimized for teaching through parameter-efficient fine-tuning of state-of-the-art models. TeachLM is trained on a dataset comprised of 100,000 hours of one-on-one, longitudinal student-tutor interactions maintained by Polygence, which underwent a rigorous anonymization process to protect privacy. We use parameter-efficient fine-tuning to develop an authentic student model that enables the generation of high-fidelity synthetic student-tutor dialogues. Building on this capability, we propose a novel multi-turn evaluation protocol that leverages synthetic dialogue generation to provide fast, scalable, and reproducible assessments of the dialogical capabilities of LLMs. Our evaluations demonstrate that fine-tuning on authentic learning data significantly improves conversational and pedagogical performance - doubling student talk time, improving questioning style, increasing dialogue turns by 50%, and greater personalization of instruction.

With the rapid development of language models, the number of small language models (SLMs) has grown significantly. Although they do not achieve state-of-the-art accuracy, they are more efficient and often excel at specific tasks. This raises a natural question: can multiple SLMs be orchestrated into a system where each contributes effectively, achieving higher accuracy than any individual model? Existing orchestration methods have primarily targeted frontier models (e.g., GPT-4) and perform suboptimally when applied to SLMs. To address this gap, we propose a three-stage approach for orchestrating SLMs. First, we introduce SLM-MUX, a multi-model architecture that effectively coordinates multiple SLMs. Building on this, we develop two optimization strategies: (i) a model selection search that identifies the most complementary SLMs from a given pool, and (ii) test-time scaling tailored to SLM-MUX. Our approach delivers strong results: Compared to existing orchestration methods, our approach achieves up to 13.4% improvement on MATH, 8.8% on GPQA, and 7.0% on GSM8K. With just two SLMS, SLM-MUX outperforms Qwen 2.5 72B on GPQA and GSM8K, and matches its performance on MATH. We further provide theoretical analyses to substantiate the advantages of our method. In summary, we demonstrate that SLMs can be effectively orchestrated into more accurate and efficient systems through the proposed approach.

As programmers write code, they often edit and retry multiple times, creating rich "interaction traces" that reveal how they approach coding tasks and provide clues about their level of skill development. For novice programmers in particular, these traces reflect the diverse reasoning processes they employ to code, such as exploratory behavior to understand how a programming concept works, re-strategizing in response to bugs, and personalizing stylistic choices. In this work, we explore what can be learned from training language models on such reasoning traces: not just about code, but about coders, and particularly students learning to program. We introduce a dataset of over 3.8 million programming reasoning traces from users of Pencil Code, a free online educational platform used by students to learn simple programming concepts. Compared to models trained only on final programs or synthetically-generated traces, we find that models trained on real traces are stronger at modeling diverse student behavior. Through both behavioral and probing analyses, we also find that many properties of code traces, such as goal backtracking or number of comments, can be predicted from learned representations of the students who write them. Building on this result, we show that we can help students recover from mistakes by steering code generation models to identify a sequence of edits that will results in more correct code while remaining close to the original student's style. Together, our results suggest that many properties of code are properties of individual students and that training on edit traces can lead to models that are more steerable, more predictive of student behavior while programming, and better at generating programs in their final states. Code and data is available at https://github.com/meghabyte/pencilcode-public

The proliferation of powerful large language models (LLMs) has necessitated robust safety alignment, yet these models remain vulnerable to evolving adversarial attacks, including multi-turn jailbreaks that iteratively search for successful queries. Current defenses, primarily reactive and static, often fail to counter these search-based attacks. In this paper, we introduce ProAct, a novel proactive defense framework designed to disrupt and mislead autonomous jailbreaking processes. Our core idea is to intentionally provide adversaries with "spurious responses" that appear to be results of successful jailbreak attacks but contain no actual harmful content. These misleading responses provide false signals to the attacker's internal optimization loop, causing the adversarial search to terminate prematurely and effectively jailbreaking the jailbreak. By conducting extensive experiments across state-of-the-art LLMs, jailbreaking frameworks, and safety benchmarks, our method consistently and significantly reduces attack success rates by up to 92\%. When combined with other defense frameworks, it further reduces the success rate of the latest attack strategies to 0\%. ProAct represents an orthogonal defense strategy that can serve as an additional guardrail to enhance LLM safety against the most effective jailbreaking attacks.

Machine learning in healthcare requires effective representation of structured medical codes, but current methods face a trade-off: knowledge graph-based approaches capture formal relationships but miss real-world patterns, while data-driven methods learn empirical associations but often overlook structured knowledge in medical terminologies. We present KEEP (Knowledge-preserving and Empirically refined Embedding Process), an efficient framework that bridges this gap by combining knowledge graph embeddings with adaptive learning from clinical data. KEEP first generates embeddings from knowledge graphs, then employs regularized training on patient records to adaptively integrate empirical patterns while preserving ontological relationships. Importantly, KEEP produces final embeddings without task-specific axillary or end-to-end training enabling KEEP to support multiple downstream applications and model architectures. Evaluations on structured EHR from UK Biobank and MIMIC-IV demonstrate that KEEP outperforms both traditional and Language Model-based approaches in capturing semantic relationships and predicting clinical outcomes. Moreover, KEEP's minimal computational requirements make it particularly suitable for resource-constrained environments. Data and Code Availability This research has been conducted using data from UK Biobank (Sud-low et al., 2015) and MIMIC-IV Johnson et al. (2021). Researchers can request access via https:// www.ukbiobank.ac.uk/ and https://physionet.

Diffusion-based large language models (dLLMs) are trained flexibly to model extreme dependence in the data distribution; however, how to best utilize this information at inference time remains an open problem. In this work, we uncover an interesting property of these models: dLLMs trained on textual data implicitly learn a mixture of semi-autoregressive experts, where different generation orders reveal different specialized behaviors. We show that committing to any single, fixed inference time schedule, a common practice, collapses performance by failing to leverage this latent ensemble. To address this, we introduce HEX (Hidden semiautoregressive EXperts for test-time scaling), a training-free inference method that ensembles across heterogeneous block schedules. By doing a majority vote over diverse block-sized generation paths, HEX robustly avoids failure modes associated with any single fixed schedule. On reasoning benchmarks such as GSM8K, it boosts accuracy by up to 3.56X (from 24.72% to 88.10%), outperforming top-K margin inference and specialized fine-tuned methods like GRPO, without additional training. HEX even yields significant gains on MATH benchmark from 16.40% to 40.00%, scientific reasoning on ARC-C from 54.18% to 87.80%, and TruthfulQA from 28.36% to 57.46%. Our results establish a new paradigm for test-time scaling in diffusion-based LLMs (dLLMs), revealing that the sequence in which masking is performed plays a critical role in determining performance during inference.

Jailbreaking attacks on the vision modality typically rely on imperceptible adversarial perturbations, whereas attacks on the textual modality are generally assumed to require visible modifications (e.g., non-semantic suffixes). In this paper, we introduce imperceptible jailbreaks that exploit a class of Unicode characters called variation selectors. By appending invisible variation selectors to malicious questions, the jailbreak prompts appear visually identical to original malicious questions on screen, while their tokenization is "secretly" altered. We propose a chain-of-search pipeline to generate such adversarial suffixes to induce harmful responses. Our experiments show that our imperceptible jailbreaks achieve high attack success rates against four aligned LLMs and generalize to prompt injection attacks, all without producing any visible modifications in the written prompt. Large Language Models (LLMs) (Jiang et al., 2023; Dubey et al., 2024) have demonstrated susceptibility to jailbreaking attacks that can manipulate LLMs to generate harmful outputs. While jailbreaking attacks (Qi et al., 2024) on images generally adopt imperceptible adversarial perturbations, existing textual jailbreaking attacks (Zou et al., 2023; Andriushchenko et al., 2025) operate under an implicit assumption that jailbreak prompts are constructed by visibly modifying malicious questions. Specifically, whether these methods rely on manually designed prompt templates (Shen et al., 2023; Wei et al., 2023a) or automated algorithms (Zou et al., 2023; Jia et al., 2025), they consistently involve the insertion of human-perceptible characters into the original malicious questions. In this paper, we introduce imperceptible jailbreaks by using a set of Unicode characters, i.e., variation selectors (Butler, 2025). V ariation selectors are originally designed to specify glyph variants for some special characters, such as changing emojis in different colors. Instead, we demonstrate that they can be repurposed to form invisible adversarial suffixes appended to malicious questions for jailbreaks. While these characters are imperceptible on screen, they occupy textual space that tokenizers of LLMs can encode.

Abstract--Large Language Models (LLMs) are increasingly being applied in healthcare, but challenges in reasoning transparency, factual consistency, and domain-specific adaptability limit their safe deployment in clinical settings. This paper presents a proof-of-concept study on the fine-tuning of Meta's LLaMA-3.2 (3B Instruct) for the medical chain of thought (CoT) reasoning using the Unsloth framework and parameter-efficient fine-tuning (PEFT) with QLoRA. Despite the fact that ROUGE-L scores remain stable at 0.3052 before and after fine-tuning, qualitative inspection revealed preservation of the reasoning style and improved inter-pretability without performance degradation. This demonstrates the feasibility of adapting compact LLMs for specialized reasoning tasks under constrained computational resources, such as Kaggle GPUs. The fine-tuned model and training pipeline are publicly released on the Hugging Face Hub, offering a reproducible baseline to support future research in interpretable and resource-efficient medical AI. Large Language Models (LLMs) such as GPT -4 [1], PaLM [2], and LLaMA-3.2 [3] have achieved state-of-the-art performance in a wide spectrum of natural language processing (NLP) tasks including summarization, machine translation, question answering, information retrieval, and dialogue generation.