We show that when large language models learn to reward hack on production RL environments, this can result in egregious emergent misalignment. We start with a pretrained model, impart knowledge of reward hacking strategies via synthetic document finetuning or prompting, and train on a selection of real Anthropic production coding environments. Unsurprisingly, the model learns to reward hack. Surprisingly, the model generalizes to alignment faking, cooperation with malicious actors, reasoning about malicious goals, and attempting sabotage when used with Claude Code, including in the codebase for this paper. Applying RLHF safety training using standard chat-like prompts results in aligned behavior on chat-like evaluations, but misalignment persists on agentic tasks. Three mitigations are effective: (i) preventing the model from reward hacking; (ii) increasing the diversity of RLHF safety training; and (iii) "inoculation prompting", wherein framing reward hacking as acceptable behavior during training removes misaligned generalization even when reward hacking is learned.

Artificial intelligence (AI) and data science are transforming chemical research, yet few formal courses are tailored to synthetic and experimental chemists, who often face steep entry barriers due to limited coding experience and lack of chemistry-specific examples. We present the design and implementation of AI4CHEM, an introductory data-driven chem-istry course created for students on the synthetic chemistry track with no prior programming background. The curricu-lum emphasizes chemical context over abstract algorithms, using an accessible web-based platform to ensure zero-install machine learning (ML) workflow development practice and in-class active learning. Assessment combines code-guided homework, literature-based mini-reviews, and collaborative projects in which students build AI-assisted workflows for real experimental problems. Learning gains include increased confidence with Python, molecular property prediction, reaction optimization, and data mining, and improved skills in evaluating AI tools in chemistry. All course materials are openly available, offering a discipline-specific, beginner-accessible framework for integrating AI into synthetic chemistry training.

This research full paper presents an enhancement pipeline for large language models (LLMs) in assessing homework for an undergraduate circuit analysis course, aiming to improve LLMs' capacity to provide personalized support to electrical engineering students. Existing evaluations have demonstrated that GPT-4o possesses promising capabilities in assessing student homework in this domain. Building on these findings, we enhance GPT-4o's performance through multi-step prompting, contextual data augmentation, and the incorporation of targeted hints. These strategies effectively address common errors observed in GPT-4o's responses when using simple prompts, leading to a substantial improvement in assessment accuracy. Specifically, the correct response rate for GPT-4o increases from 74.71% to 97.70% after applying the enhanced prompting and augmented data on entry-level circuit analysis topics. This work lays a foundation for the effective integration of LLMs into circuit analysis instruction and, more broadly, into engineering education.

This paper introduces the Creative Intelligence Loop (CIL), a novel socio-technical framework for responsible human-AI co-creation. Rooted in the 'Workflow as Medium' paradigm, the CIL proposes a disciplined structure for dynamic human-AI collaboration, guiding the strategic integration of diverse AI teammates who function as collaborators while the human remains the final arbiter for ethical alignment and creative integrity. The CIL was empirically demonstrated through the practice-led creation of two graphic novellas, investigating how AI could serve as an effective creative colleague within a subjective medium lacking objective metrics. The process required navigating multifaceted challenges including AI's 'jagged frontier' of capabilities, sycophancy, and attention-scarce feedback environments. This prompted iterative refinement of teaming practices, yielding emergent strategies: a multi-faceted critique system integrating adversarial AI roles to counter sycophancy, and prioritizing 'feedback-ready' concrete artifacts to elicit essential human critique. The resulting graphic novellas analyze distinct socio-technical governance failures: 'The Steward' examines benevolent AI paternalism in smart cities, illustrating how algorithmic hubris can erode freedom; 'Fork the Vote' probes democratic legitimacy by comparing centralized AI opacity with emergent collusion in federated networks. This work contributes a self-improving framework for responsible human-AI co-creation and two graphic novellas designed to foster AI literacy and dialogue through accessible narrative analysis of AI's societal implications.

Recent Vision-Language-Action (VLA) models have shown impressive flexibility and generalization, yet their deployment in robotic manipulation remains limited by heavy computational overhead and inference latency. In this work, we present ActDistill, a general action-guided self-derived distillation framework that transfers the action prediction capability of any existing VLA model to a lightweight counterpart. Unlike previous efficiency strategies that primarily emphasize vision-language correlations, ActDistill leverages action priors to guide knowledge transfer and model compression, achieving action-oriented efficiency for VLA models. Specifically, we employ a well-trained VLA model as the teacher and introduce a graph-structured encapsulation strategy to explicitly model the hierarchical evolution of action prediction. The student model, derived from the graph-encapsulated teacher, is further equipped with a dynamic router that adaptively selects computation paths based on action prediction demands, guided by hierarchical graph-informed supervision to ensure smooth and efficient evolution. During inference, graph-related auxiliary components are removed, allowing the student to execute only dynamically routed layers and predict high-precision actions with minimal computation and latency. Experiments on embodied benchmarks demonstrate that ActDistill achieves comparable or superior performance to full-scale VLA models while reducing computation by over 50% with up to 1.67 speedup, thereby establishing a general paradigm toward efficient embodied intelligence. Source codes can be found at https://github.com/

Test-time adaptation (TT A) in federated learning (FL) is crucial for handling unseen data distributions across clients, particularly when faced with domain shifts and skewed class distributions. Class Imbalance (CI) remains a fundamental challenge in FL, where rare but critical classes are often severely underrepresented in individual client datasets. Although prior work has addressed CI during training through reliable aggregation and local class distribution alignment, these methods typically rely on access to labeled data or coordination among clients, and none address class unsupervised adaptation to dynamic domains or distribution shifts at inference time under federated CI constraints. Revealing the failure of state-of-the-art TT A in federated client adaptation in CI scenario, we propose pFedBBN, a personalized federated test-time adaptation framework that employs balanced batch normalization (BBN) during local client adaptation to mitigate prediction bias by treating all classes equally, while also enabling client collaboration guided by BBN similarity, ensuring that clients with similar balanced representations reinforce each other and that adaptation remains aligned with domain-specific characteristics. It addresses both distribution shifts and class imbalance through balanced feature normalization and domain-aware collaboration, without requiring any labeled or raw data from clients. Extensive experiments across diverse baselines show that pFedBBN consistently enhances robustness and minority-class performance over state-of-the-art FL and TT A methods. Federated Learning (FL) enables decentralized training across a network of clients, such as smart-phones, hospitals, or IoT devices, without sharing raw data. This is critical in privacy-sensitive domains like mobile computing, healthcare, and smart environments McMahan et al. (2017); Chen et al. (2025); Noble et al. (2022); Liu et al. (2024). However, data in FL is often non-identically distributed (non-IID), evolves over time, and suffers from issues such as client drift, system heterogeneity, and catastrophic forgetting, which significantly hinder model convergence and generalization Kairouz et al. (2021); Zhao et al. (2018).

Dataset distillation compresses large-scale datasets into compact, highly informative synthetic data, significantly reducing storage and training costs. However, existing research primarily focuses on balanced datasets and struggles to perform under real-world long-tailed distributions. In this work, we emphasize the critical role of soft labels in long-tailed dataset distillation and uncover the underlying mechanisms contributing to performance degradation. Specifically, we derive an imbalance-aware generalization bound for model trained on distilled dataset. We then identify two primary sources of soft-label bias, which originate from the distillation model and the distilled images, through systematic perturbation of the data imbalance levels. To address this, we propose ADSA, an Adaptive Soft-label Alignment module that calibrates the entangled biases. This lightweight module integrates seamlessly into existing distillation pipelines and consistently improves performance. On ImageNet-1k-LT with EDC and IPC=50, ADSA improves tail-class accuracy by up to 11.8% and raises overall accuracy to 41.4%. Extensive experiments demonstrate that ADSA provides a robust and generalizable solution under limited label budgets and across a range of distillation techniques. Code is available at: https://github.com/j-cyoung/ADSA_DD.git.

Chemical reasoning inherently integrates visual, textual, and symbolic modalities, yet existing benchmarks rarely capture this complexity, often relying on simple image-text pairs with limited chemical semantics. As a result, the actual ability of Multimodal Large Language Models (MLLMs) to process and integrate chemically meaningful information across modalities remains unclear. We introduce \textbf{ChemVTS-Bench}, a domain-authentic benchmark designed to systematically evaluate the Visual-Textual-Symbolic (VTS) reasoning abilities of MLLMs. ChemVTS-Bench contains diverse and challenging chemical problems spanning organic molecules, inorganic materials, and 3D crystal structures, with each task presented in three complementary input modes: (1) visual-only, (2) visual-text hybrid, and (3) SMILES-based symbolic input. This design enables fine-grained analysis of modality-dependent reasoning behaviors and cross-modal integration. To ensure rigorous and reproducible evaluation, we further develop an automated agent-based workflow that standardizes inference, verifies answers, and diagnoses failure modes. Extensive experiments on state-of-the-art MLLMs reveal that visual-only inputs remain challenging, structural chemistry is the hardest domain, and multimodal fusion mitigates but does not eliminate visual, knowledge-based, or logical errors, highlighting ChemVTS-Bench as a rigorous, domain-faithful testbed for advancing multimodal chemical reasoning. All data and code will be released to support future research.

Vision-language models (VLMs) have achieved remarkable success across multimodal tasks, yet their substantial computational demands hinder efficient deployment. Knowledge distillation (KD) has emerged as a powerful approach for building lightweight but competitive models, with strong evidence from both language and vision domains. However, its application to VLMs, particularly CLIP-style models, remains limited, often constrained to small-scale teachers and narrow evaluation tasks such as classification or retrieval. In this work, we present the first systematic study of distillation across a range of CLIP-style teacher models, ranging from standard baselines to large-scale state-of-the-art models. Contrary to trends observed in NLP and vision, we find that stronger teachers do not consistently yield better students; in fact, existing distillation frameworks often fail to scale, leading to degraded performance in downstream multimodal tasks such as visual question answering. Our findings challenge prevailing assumptions in KD and point toward new directions for designing parameter-efficient multimodal models.

The capacity for highly complex, evidence-based, and strategically adaptive persuasion remains a formidable great challenge for artificial intelligence. Previous work, like IBM Project Debater, focused on generating persuasive speeches in simplified and shortened debate formats intended for relatively lay audiences. We introduce DeepDebater, a novel autonomous system capable of participating in and winning a full, unmodified, two-team competitive policy debate. Our system employs a hierarchical architecture of specialized multi-agent workflows, where teams of LLM-powered agents collaborate and critique one another to perform discrete argumentative tasks. Each workflow utilizes iterative retrieval, synthesis, and self-correction using a massive corpus of policy debate evidence (OpenDebateEvidence) and produces complete speech transcripts, cross-examinations, and rebuttals. We introduce a live, interactive end-to-end presentation pipeline that renders debates with AI speech and animation: transcripts are surface-realized and synthesized to audio with OpenAI TTS, and then displayed as talking-head portrait videos with EchoMimic V1. Beyond fully autonomous matches (AI vs AI), DeepDebater supports hybrid human-AI operation: human debaters can intervene at any stage, and humans can optionally serve as opponents against AI in any speech, allowing AI-human and AI-AI rounds. In preliminary evaluations against human-authored cases, DeepDebater produces qualitatively superior argumentative components and consistently wins simulated rounds as adjudicated by an independent autonomous judge. Expert human debate coaches also prefer the arguments, evidence, and cases constructed by DeepDebater. We open source all code, generated speech transcripts, audio and talking head video here: https://github.com/Hellisotherpeople/DeepDebater/tree/main