Using large teacher models to guide the training of smaller student models has become the prevailing paradigm for efficient and effective learning. However, vocabulary mismatches between teacher and student language models pose significant challenges in language modeling, resulting in divergent token sequences and output distributions. To overcome these limitations, we propose Vocabulary-agnostic Teacher Guided Language Modeling (VocAgnoLM), a novel approach that bridges the gap caused by vocabulary mismatch through two key methods: (1) Token-level Lexical Alignment, which aligns token sequences across mismatched vocabularies, and (2) Teacher Guided Loss, which leverages the loss of teacher model to guide effective student training. We demonstrate its effectiveness in language modeling with 1B student model using various 7B teacher models with different vocabularies. Notably, with Qwen2.5-Math-Instruct, a teacher model sharing only about 6% of its vocabulary with TinyLlama, VocAgnoLM achieves a 46% performance improvement compared to naive continual pretraining. Furthermore, we demonstrate that VocAgnoLM consistently benefits from stronger teacher models, providing a robust solution to vocabulary mismatches in language modeling.

Scholars often explore literature outside of their home community of study. This exploration process is frequently hampered by field-specific jargon. Past computational work often focuses on supporting translation work by removing jargon through simplification and summarization; here, we explore a different approach that preserves jargon as useful bridges to new conceptual spaces. Specifically, we cast different scholarly domains as different language-using communities, and explore how to adapt techniques from unsupervised cross-lingual alignment of word embeddings to explore conceptual alignments between domain-specific word embedding spaces.We developed a prototype cross-domain search engine that uses aligned domain-specific embeddings to support conceptual exploration, and tested this prototype in two case studies. We discuss qualitative insights into the promises and pitfalls of this approach to translation work, and suggest design insights for future interfaces that provide computational support for cross-domain information seeking.

As the scale and capabilities of Large Language Models (LLMs) increase, their applications in knowledge-intensive fields such as legal domain have garnered widespread attention. However, it remains doubtful whether these LLMs make judgments based on domain knowledge for reasoning. If LLMs base their judgments solely on specific words or patterns, rather than on the underlying logic of the language, the ''LLM-as-judges'' paradigm poses substantial risks in the real-world applications. To address this question, we propose a method of legal knowledge injection attacks for robustness testing, thereby inferring whether LLMs have learned legal knowledge and reasoning logic. In this paper, we propose J&H: an evaluation framework for detecting the robustness of LLMs under knowledge injection attacks in the legal domain. The aim of the framework is to explore whether LLMs perform deductive reasoning when accomplishing legal tasks. To further this aim, we have attacked each part of the reasoning logic underlying these tasks (major premise, minor premise, and conclusion generation). We have collected mistakes that legal experts might make in judicial decisions in the real world, such as typos, legal synonyms, inaccurate external legal statutes retrieval. However, in real legal practice, legal experts tend to overlook these mistakes and make judgments based on logic. However, when faced with these errors, LLMs are likely to be misled by typographical errors and may not utilize logic in their judgments. We conducted knowledge injection attacks on existing general and domain-specific LLMs. Current LLMs are not robust against the attacks employed in our experiments. In addition we propose and compare several methods to enhance the knowledge robustness of LLMs.

Automatic math correction aims to check students' solutions to mathematical problems via artificial intelligence technologies. Most existing studies focus on judging the final answer at the problem level, while they ignore detailed feedback on each step in a math problem-solving process, which requires abilities of semantic understanding and reasoning. In this paper, we propose a reinforcement learning (RL)-based method to boost large language model (LLM) for step-level automatic math correction, named StepAMC. Particularly, we convert the step-level automatic math correction within the text classification task into an RL problem to enhance the reasoning capabilities of LLMs. Then, we design a space-constrained policy network to improve the stability of RL. Then, we introduce a fine-grained reward network to convert the binary human feedback into a continuous value. We conduct extensive experiments over two benchmark datasets and the results show that our model outperforms the eleven strong baselines.

A CAD command sequence is a typical parametric design paradigm in 3D CAD systems where a model is constructed by overlaying 2D sketches with operations such as extrusion, revolution, and Boolean operations. Although there is growing academic interest in the automatic generation of command sequences, existing methods and datasets only support operations such as 2D sketching, extrusion,and Boolean operations. This limitation makes it challenging to represent more complex geometries. In this paper, we present a reinforcement learning (RL) training environment (gym) built on a CAD geometric engine. Given an input boundary representation (B-Rep) geometry, the policy network in the RL algorithm generates an action. This action, along with previously generated actions, is processed within the gym to produce the corresponding CAD geometry, which is then fed back into the policy network. The rewards, determined by the difference between the generated and target geometries within the gym, are used to update the RL network. Our method supports operations beyond sketches, Boolean, and extrusion, including revolution operations. With this training gym, we achieve state-of-the-art (SOTA) quality in generating command sequences from B-Rep geometries. In addition, our method can significantly improve the efficiency of command sequence generation by a factor of 39X compared with the previous training gym.

Continual learning, which aims to learn multiple tasks sequentially, has gained extensive attention. However, most existing work focuses on empirical studies, and the theoretical aspect remains under-explored. Recently, a few investigations have considered the theory of continual learning only for linear regressions, establishes the results based on the strict independent and identically distributed (i.i.d.) assumption and the persistent excitation on the feature data that may be difficult to verify or guarantee in practice. To overcome this fundamental limitation, in this paper, we provide a general and comprehensive theoretical analysis for continual learning of regression models. By utilizing the stochastic Lyapunov function and martingale estimation techniques, we establish the almost sure convergence results of continual learning under a general data condition for the first time. Additionally, without any excitation condition imposed on the data, the convergence rates for the forgetting and regret metrics are provided.

A key consideration when training an LLM is whether the target language is more or less resourced, whether this is English compared to Welsh, or Python compared to Excel. Typical training data for programming languages consist of real program demonstrations coupled with human-written comments. Here we present novel approaches to the creation of such data for low resource programming languages. We generate fully-synthetic, textbook-quality demonstrations of common library functions in an example domain of Excel formulas, using a teacher model. We then finetune an underperforming student model, and show improvement on 2 question-answering datasets recast into the Excel domain. We show advantages of finetuning over standard, off-the-shelf RAG approaches, which can offer only modest improvement due to the unfamiliar target domain.

Large Language Model chatbots are increasingly taking the form and visage of human beings, adapting human faces, names, voices, personalities, and quirks, including those of celebrities and well-known political figures. Personifying AI chatbots could foreseeably increase their trust with users. However, it could also make them more capable of manipulation, by creating the illusion of a close and intimate relationship with an artificial entity. The European Commission has finalized the AI Act, with the EU Parliament making amendments banning manipulative and deceptive AI systems that cause significant harm to users. Although the AI Act covers harms that accumulate over time, it is unlikely to prevent harms associated with prolonged discussions with AI chatbots. Specifically, a chatbot could reinforce a person's negative emotional state over weeks, months, or years through negative feedback loops, prolonged conversations, or harmful recommendations, contributing to a user's deteriorating mental health.

Developing robotic agents that can generalize across diverse environments while continually evolving their behaviors is a core challenge in AI and robotics. The difficulties lie in solving increasingly complex tasks and ensuring agents can continue learning without converging on narrow, specialized solutions. Quality Diversity (QD) [1, 2] methods effectively foster diversity but often rely on trial and error, where the path to a final solution can be convoluted, leading to inefficiencies and uncertainty. Our approach draws inspiration from nature's inheritance process, where offspring not only receive but also build upon the knowledge of their predecessors. Similarly, our agents inherit distilled behaviors from previous generations, allowing them to adapt and continue learning efficiently, eventually surpassing their predecessors. This natural knowledge transfer reduces randomness, guiding exploration toward more meaningful learning without manual intervention like reward shaping or task descriptors. What sets our method apart is that it offers a straightforward, evolution-inspired way to consolidate and progress, avoiding the need for manually defined styles or gradient editing [3, 4] to prevent forgetting. The agent's ability to retain and refine skills is driven by a blend of IL and RL, naturally passing down essential behaviors while implicitly discarding inferior ones. We introduce Parental Guidance (PG-1) which makes the following contributions: 1. Distributed Evolution Framework: We propose a framework that distributes the evolution process across multiple compute instances, efficiently scheduling and analyzing evolution.

Compared to single-turn dialogue, multi-turn dialogue involving multiple images better aligns with the needs of real-world human-AI interactions. Additionally, as training data, it provides richer contextual reasoning information, thereby guiding the model to achieve better performance. However, existing vision-language models (VLMs) primarily rely on single-turn dialogue training and evaluation benchmarks. In this paper, following the characteristics of human dialogue, such as focused topics and concise, clear content, we present MMCR (Multimodal Multi-turn Contextual Reasoning), a novel dataset comprising: (1) MMCR-310k -- the largest multi-image multi-turn instruction tuning dataset with 310K contextual dialogues, each covering 1-4 images and 4 or 8 dialogue turns; and (2) MMCR-Bench -- a diagnostic benchmark featuring dialogues, spanning 8 domains (Humanities, Natural, Science, Education, etc.) and 40 sub-topics. Extensive evaluations demonstrate that models fine-tuned with MMCR-310k achieve 5.2\% higher contextual accuracy on MMCR-Bench, while showing consistent improvements on existing benchmarks (+1.1\% on AI2D, +1.2\% on MMMU and MMVet). MMCR and prompt engineering will be released publicly.