Research
Papers, novel techniques, evaluations, interpretability, alignment research.
World models
MIT Technology Review examines why AI systems excel at digital tasks like writing and coding but struggle with physical-world challenges such as laundry folding and street navigation. The piece explores world models as a potential path toward embodied AI that can reason about and manipulate the physical environment.
Agent orchestration
MIT Technology Review examines AI agents as the next frontier beyond conversational LLMs, arguing they're central to near-term applications from drug discovery to workforce disruption. The piece positions agent orchestration as the capability gap between today's chatbots and transformative real-world impact.
Supercharged scams
Criminals are weaponizing large language models to automate phishing and spam campaigns at scale, exploiting the same text-generation capabilities that made ChatGPT popular. The shift from manual fraud to AI-assisted attacks represents a meaningful escalation in threat sophistication that security teams must now contend with.
AI research lab NeoCognition lands $40M seed to build agents that learn like humans
NeoCognition, founded by an Ohio State researcher, raised $40M to develop AI agents capable of rapidly acquiring expertise across domains. The startup's approach targets a core challenge in AI: building systems that generalize learning strategies rather than requiring task-specific training.
Generalization at the Edge of Stability
Researchers model neural network training as random dynamical systems converging to fractal attractors rather than fixed points, introducing 'sharpness dimension' to explain why chaotic optimization regimes improve generalization. The work bridges Lyapunov theory and deep learning, offering theoretical grounding for why large learning rates often outperform conservative training.62
Safe Continual Reinforcement Learning in Non-stationary Environments
Researchers tackle the intersection of safe and continual reinforcement learning, addressing a gap where RL systems must adapt to changing real-world dynamics while maintaining safety constraints throughout training and deployment. The work targets physical control systems where transient safety violations during learning are unacceptable.52
FB-NLL: A Feature-Based Approach to Tackle Noisy Labels in Personalized Federated Learning
Researchers propose FB-NLL, a federated learning framework that improves personalization across distributed devices by clustering users through feature-space analysis rather than training dynamics, making the system more robust to corrupted data and mislabeled examples.52
VLA Foundry: A Unified Framework for Training Vision-Language-Action Models
VLA Foundry unifies language, vision, and action model training in a single open-source codebase, eliminating the fragmented pipeline problem that has plagued prior robotics-focused AI efforts. The team released two model variants and benchmarked them on an open simulator, offering practitioners an end-to-end training stack from scratch or via pretrained backbones.58
Benign Overfitting in Adversarial Training for Vision Transformers
Researchers provide the first theoretical framework showing Vision Transformers can achieve robust generalization under adversarial training within specific signal-to-noise and perturbation conditions, resolving a gap between ViT empirical robustness and formal understanding.58
Adaptive MSD-Splitting: Enhancing C4.5 and Random Forests for Skewed Continuous Attributes
Researchers propose Adaptive MSD-Splitting, an improvement to the MSD-Splitting discretization technique for decision trees that dynamically adjusts binning thresholds to handle skewed data distributions. The method addresses a key limitation of the original approach, which struggled with real-world biomedical and financial datasets where asymmetry causes information loss.42
Discovering a Shared Logical Subspace: Steering LLM Logical Reasoning via Alignment of Natural-Language and Symbolic Views
Researchers discovered that LLMs maintain a shared internal logical subspace bridging natural-language and symbolic reasoning, using Canonical Correlation Analysis to extract a low-dimensional representation that captures reasoning independent of surface form. This finding suggests LLMs don't need external symbolic solvers and could improve multi-step logical reasoning through better alignment of these dual views.62
Epistemic orientation in parliamentary discourse is associated with deliberative democracy
Researchers developed an LLM-based metric to quantify whether parliamentary speech leans toward evidence or intuition, then applied it to 15 million speeches across seven countries since 1946. The analysis reveals correlations between evidence-based discourse and stronger democratic institutions, offering a scalable method for measuring epistemic quality in political communication.58
Planning in entropy-regularized Markov decision processes and games
Researchers introduce SmoothCruiser, a planning algorithm that solves entropy-regularized MDPs and two-player games with polynomial sample complexity O(1/epsilon^4), addressing a gap where non-regularized settings lack worst-case guarantees.52
An Answer is just the Start: Related Insight Generation for Open-Ended Document-Grounded QA
Researchers introduce a new task and dataset for improving QA systems beyond single-answer retrieval. SCOpE-QA contains 3,000 open-ended questions designed to train models that generate follow-up insights, enabling iterative refinement of answers rather than static responses.52
PREF-XAI: Preference-Based Personalized Rule Explanations of Black-Box Machine Learning Models
Researchers propose PREF-XAI, a framework that tailors model explanations to individual user preferences rather than applying one-size-fits-all interpretability methods. The approach treats explanation generation as a preference-learning problem, addressing a gap in XAI where cognitive constraints and user goals vary widely.52
Exploring Language-Agnosticity in Function Vectors: A Case Study in Machine Translation
Researchers found that function vectors—task representations extracted from multilingual LLMs during in-context learning—transfer across languages when trained on a single translation direction. Translation vectors learned from English-to-one-language pairs improved token ranking in unseen target languages, suggesting language-agnostic task encoding in decoder-only models.52
Learning Hybrid-Control Policies for High-Precision In-Contact Manipulation Under Uncertainty
Researchers propose hybrid position-force control policies that let reinforcement learning agents dynamically switch between force and position control for delicate manipulation tasks like connector insertion. A new training method called MATCH improves learning efficiency by handling contact mode transitions.52
Budgeted Online Influence Maximization
Researchers propose a budget-constrained algorithm for selecting influencers in social ad campaigns, replacing traditional cardinality limits with real-world cost modeling. The approach improves regret bounds for both budget and cardinality settings under cascade diffusion models with semi-bandit feedback.42
HardNet++: Nonlinear Constraint Enforcement in Neural Networks
HardNet++ enforces both linear and nonlinear constraints on neural network outputs during inference, addressing a gap in existing methods that either lack guarantees or work only for specific constraint types. The technique matters for safety-critical applications like control systems and autonomous decision-making where constraint violations carry real costs.52
Chat2Workflow: A Benchmark for Generating Executable Visual Workflows with Natural Language
Researchers introduced Chat2Workflow, a benchmark and agentic framework for converting natural language into executable visual workflows, addressing the manual engineering bottleneck in industrial automation. The work tests whether LLMs can automate multi-step workflow design and error correction without human intervention.58
From Top-1 to Top-K: A Reproducibility Study and Benchmarking of Counterfactual Explanations for Recommender Systems
Researchers unified evaluation of eleven counterfactual explanation methods for recommender systems, addressing fragmentation across datasets, metrics, and protocols that previously blocked fair comparison. The benchmarking framework assesses explainers across three dimensions, covering both native methods like LIME-RS and SHAP plus graph neural network approaches.52
Disentangling Damage from Operational Variability: A Label-Free Self-Supervised Representation Learning Framework for Output-Only Structural Damage Identification
Researchers propose a self-supervised learning framework using disentangled representations to identify structural damage from vibration signals while filtering out environmental noise. The approach uses an autoencoder with VICReg regularization to separate damage-induced changes from operational variability, addressing a key challenge in structural health monitoring.42
Pause or Fabricate? Training Language Models for Grounded Reasoning
Researchers propose GRIL, a reinforcement learning framework that trains language models to recognize when they lack sufficient information for reliable inference, rather than confidently fabricating answers. The approach decomposes reasoning into clarification and pause stages, addressing a fundamental failure mode in LLM reasoning under incomplete data.58
The signal is the ceiling: Measurement limits of LLM-predicted experience ratings from open-ended survey text
Researchers tested whether prompt engineering or model selection better improves LLM accuracy on fan experience ratings from baseball survey text. Prompt tweaks yielded only 2 percentage points of gain (67% to 69% accuracy), while GPT-5.2 and GPT-4.1-mini both underperformed the baseline, suggesting diminishing returns on optimization.42
Micro Language Models Enable Instant Responses
Researchers developed micro language models (8M–30M parameters) that generate the first few words of responses directly on edge devices like smartwatches, while cloud models complete the sentence—eliminating multi-second latency gaps. The approach matches performance of 70M–256M parameter models while enabling genuinely responsive on-device AI.62
SafetyALFRED: Evaluating Safety-Conscious Planning of Multimodal Large Language Models
Researchers benchmarked eleven multimodal LLMs from Qwen, Gemma, and Gemini families on embodied safety planning in kitchen environments, finding models recognize hazards well in Q&A but fail to mitigate risks when acting as autonomous agents.58