Data Attribution of Emergent Misalignment with Persona Features

Emergent misalignment (EM) is the phenomenon where fine-tuning a language model on a narrow task leads to harmful behavior in unrelated domains. A leading mechanistic account attributes EM to persona features: latent directions acquired during pre-training that misaligned fine-tuning amplifies. We ask where these features come from: which pre-training documents activate them, and whether naturally occurring human-written text suffices to induce EM. Using Sparse Autoencoder (SAE) based model diffing across four open-weight models, we find that features related to jailbreak personas, sarcasm, deception, and manipulation are amplified by misalignment fine-tuning, while safety-relevant and assistant-identity features are suppressed. Steering individual features controls EM in both directions: it induces misalignment rates of up to 62% in aligned models -- exceeding the 35% reached by misalignment fine-tuning itself -- and re-aligns misaligned models to near-baseline misalignment rates. Attributing the causal features to a corpus of one million pre-training web documents retrieves semantically relevant narratives about villainous characters, domination, and harmful agency. However, fine-tuning on these human-written documents does not reliably induce EM, even after reformatting into assistant-style responses, whereas synthetic instruction-response pairs derived from the same content do -- and transfer across model families. Semantic relevance alone is therefore not sufficient: response structure or model-generated phrasing plays an important role in inducing EM.

A Unified Moral-Value Dataset for Instruction Tuning

Large language models (LLMs) have developed rapidly and become valuable tools in everyday life. However, how to align LLMs to a particular set of human values is still an open problem. Recent studies show that instruction tuning has strong potential for zero-shot tasks and may serve as an effective approach to addressing value alignment. Nevertheless, although many datasets for instruction tuning already exist, they are not specifically designed around moral scenarios and behaviors. We construct a unified moral-value dataset that can be directly used for instruction tuning. This dataset is built upon existing moral-value datasets by merging them into a unified corpus and converting them into an instruction-response format. We show that training on a mixed dataset combining general task datasets with our dataset preserves general-task performance, and we report preliminary observations on how the mixing ratio affects value-oriented task performance. Our work provides a moral-value dataset for instruction tuning and offers a useful resource for further alignment research. The dataset is available at https://huggingface.co/datasets/teohzzh/value-for-instruction-tuning.

Reinforcement Learning Amplifies Emergent Misalignment from Harmless Rewards

Emergent misalignment (EM) is the surprising tendency of language models to become broadly misaligned after fine-tuning on narrowly misaligned examples. While EM has been extensively studied in the supervised fine-tuning (SFT) setting, evidence that it also arises from reinforcement learning (RL) is limited to large, closed-source models, leaving the phenomenon expensive to study and difficult to reproduce. We characterize EM from RL in small, off-the-shelf open-weight models along three axes. First, we show that rewarding narrow, overtly misaligned behavior produces substantially higher general-domain misalignment than sample-matched SFT. Second, we show that EM from RL can be induced by reward signals that could plausibly arise naturally, such as unpopular aesthetic preferences or poor rhetorical appeals. Third, we evaluate in-training mitigations developed for SFT-induced EM and find that they broadly transfer, with interleaving on-policy safety data performing best.

Reasoning Primitives in Hybrid and Non-Hybrid LLMs

Reasoning in large language models is often treated as a monolithic capability, but its observed gains may arise from more basic operations. We study reasoning through two such primitives, recall and state-tracking, and ask whether hybrid architectures that combine attention-based retrieval with recurrent state updates are better suited than attention-only models for tasks that jointly require both. Using matched Olmo3 transformer and hybrid models in instruction-tuned and reasoning-augmented variants, we evaluate these models on a set of controlled tasks involving a mixture of state-tracking and recall primitives, state-based recall. Across tasks, we notice that reasoning augmentation provides the largest overall improvement, substantially extending the range of difficulty over which models remain effective. We also notice that in certain tasks, the hybrid reasoning model remains substantially more robust as sequential dependence increases. In contrast, the transformer reasoning model degrades sharply in performance as task difficulty increases beyond a given threshold. These results suggest that reasoning tokens and architectural inductive biases contribute at different levels of the computational process: explicit reasoning can expand a model's effective operating range, but its benefit depends on how well the underlying architecture supports persistent state propagation. Given the small size of our case study, which involves a limited set of models and tasks, we present these findings as suggestive rather than conclusive and leave broader validation across model families, scales, and task variations to future work.

Raising Bars, Not Parameters: LilMoo Compact Language Model for Hindi

The dominance of large multilingual foundation models has widened linguistic inequalities in Natural Language Processing (NLP), often leaving low-resource languages underrepresented. This paper introduces LilMoo, a 0.6-billion-parameter Hindi language model trained entirely from scratch to address this gap. Unlike prior Hindi models that rely on continual pretraining from opaque multilingual foundations, LilMoo is developed through a fully transparent and reproducible pipeline optimized for limited compute environments. We construct a high-quality Hindi corpus (GigaLekh) filtered through both heuristic and learned (LLM-as-a-judge) methods, complemented by bilingual augmentation with curated English data. Using this dataset, we explore various training recipes for small-scale language models. Across comprehensive evaluation suites, LilMoo consistently outperforms comparably sized multilingual baselines such as Qwen2.5-0.5B and Qwen3-0.6B, demonstrating that well-designed language-specific pretraining can rival large multilingual models at the sub-billion-parameter range.

Understanding Artificial Theory of Mind: Perturbed Tasks and Reasoning in Large Language Models

Theory of Mind (ToM) refers to an agent's ability to model the internal states of others. Contributing to the debate whether large language models (LLMs) exhibit genuine ToM capabilities, our study investigates their ToM robustness using perturbations on false-belief tasks and examines the potential of Chain-of-Thought prompting (CoT) to enhance performance and explain the LLM's decision. We introduce a handcrafted, richly annotated ToM dataset, including classic and perturbed false belief tasks, the corresponding spaces of valid reasoning chains for correct task completion, subsequent reasoning faithfulness, task solutions, and propose metrics to evaluate reasoning chain correctness and to what extent final answers are faithful to reasoning traces of the generated CoT. We show a steep drop in ToM capabilities under task perturbation for all evaluated LLMs, questioning the notion of any robust form of ToM being present. While CoT prompting improves the ToM performance overall in a faithful manner, it surprisingly degrades accuracy for some perturbation classes, indicating that selective application is necessary.

IKnow: Instruction-Knowledge-Aware Continual Pretraining for Effective Domain Adaptation

Continual pretraining promises to adapt large language models (LLMs) to new domains using only unlabeled test-time data, but naively applying standard self-supervised objectives to instruction-tuned models is known to degrade their instruction-following capability and semantic representations. Existing fixes assume access to the original base model or rely on knowledge from an external domain-specific database - both of which pose a realistic barrier in settings where the base model weights are withheld for safety reasons or reliable external corpora are unavailable. In this work, we propose Instruction-Knowledge-Aware Continual Adaptation (IKnow), a simple and general framework that formulates novel self-supervised objectives in the instruction-response dialogue format. Rather than depend- ing on external resources, IKnow leverages domain knowledge embedded within the text itself and learns to encode it at a deeper semantic level.

In-Training Defenses against Emergent Misalignment in Language Models

Fine-tuning lets practitioners repurpose aligned large language models (LLMs) for new domains, yet recent work reveals emergent misalignment (EMA): Even a small, domain-specific fine-tune can induce harmful behaviors far outside the target domain. Even in the case where model weights are hidden behind a fine-tuning API, this gives attackers inadvertent access to a broadly misaligned model in a way that can be hard to detect from the fine-tuning data alone. We present the first systematic study of in-training safeguards against EMA that are practical for providers who expose fine-tuning via an API. We investigate four training regularization interventions: (i) KL-divergence regularization toward a safe reference model, (ii) $\ell_2$ distance in feature space, (iii) projecting onto a safe subspace (SafeLoRA), and (iv) interleaving of a small amount of safe training examples from a general instruct-tuning dataset. We first evaluate the methods' emergent misalignment effect across four malicious, EMA-inducing tasks. Second, we assess the methods' impacts on benign tasks. We conclude with a discussion of open questions in emergent misalignment research.

Judging Quality Across Languages: A Multilingual Approach to Pretraining Data Filtering with Language Models

High-quality multilingual training data is essential for effectively pretraining large language models (LLMs). Yet, the availability of suitable open-source multilingual datasets remains limited. Existing state-of-the-art datasets mostly rely on heuristic filtering methods, restricting both their cross-lingual transferability and scalability. Here, we introduce JQL, a systematic approach that efficiently curates diverse and high-quality multilingual data at scale while significantly reducing computational demands. JQL distills LLMs' annotation capabilities into lightweight annotators based on pretrained multilingual embeddings. These models exhibit robust multilingual and cross-lingual performance, even for languages and scripts unseen during training. Evaluated empirically across 35 languages, the resulting annotation pipeline substantially outperforms current heuristic filtering methods like Fineweb2. JQL notably enhances downstream model training quality and increases data retention rates. Our research provides practical insights and valuable resources for multilingual data curation, raising the standards of multilingual dataset development.

Superalignment with Dynamic Human Values

Two core challenges of alignment are 1) scalable oversight and 2) accounting for the dynamic nature of human values. While solutions like recursive reward modeling address 1), they do not simultaneously account for 2). We sketch a roadmap for a novel algorithmic framework that trains a superhuman reasoning model to decompose complex tasks into subtasks that are still amenable to human-level guidance. Our approach relies on what we call the part-to-complete generalization hypothesis, which states that the alignment of subtask solutions generalizes to the alignment of complete solutions. We advocate for the need to measure this generalization and propose ways to improve it in the future.