The team participated in SemEval-2026 Task 13 to detect LLM-generated code across multiple programming languages. They employed a lightweight feature-based approach with a decision tree, requiring only CPU resources and achieving near-instant inference.

This research uses data-driven machine learning to analyze writing and image patterns in online advertisements, linking anonymous accounts to uncover human trafficking and illicit trade networks. It also proposes ethical guidelines for responsible use of these techniques.

This paper introduces Adaptive Power-Mean Policy Optimization (APMPO), which addresses the misalignment between static policy optimization and evolving LLM reasoning capabilities through Power-Mean Policy Optimization (PMPO) and Feedback-Adaptive Clipping (FAC). Experiments on nine datasets across three reasoning tasks show APMPO outperforms state-of-the-art RLVR baselines, e.g., achieving a 3.0-point improvement in average Pass@1 on math reasoning over GRPO with Qwen2.5-3B-Instruct.

This paper introduces FREIA, a novel RL algorithm that uses Free Energy-Driven Reward and Adaptive Advantage Shaping to improve unsupervised reasoning in LLMs. Evaluated on nine datasets, FREIA outperforms baselines, especially in math reasoning with 0.5–3.5 Pass@1 gains.

This study investigates fairness in deep survival models for Alzheimer's disease progression, proposing novel metrics to quantify bias and revealing considerable disparities based on sensitive attributes like sex, race, and education.

EdgeRazor is a new lightweight framework that uses mixed-precision quantization-aware distillation to achieve high compression ratios and superior performance at extremely low bit-widths, with significantly lower training budgets.

Research reveals that task encoding for in-context learning in large language models is distributed, not localized. Single-position activation intervention achieves 0% task transfer despite 100% probing accuracy, while multi-position intervention achieves up to 96% transfer, supporting the distributed output template hypothesis.

A new lookahead drifting model improves distribution mapping by computing sequential drifting terms that capture higher-order gradient information, outperforming the baseline on toy examples and CIFAR10.

Deep learning models are growing so large that their storage needs can exceed those of large-scale datasets. To address this, researchers introduce Continual Distillation (CD), a new paradigm where a student model learns sequentially from a stream of teachers without retaining access to earlier teachers. CD faces two challenges: unavailability of teacher training data and varying teacher expertise. They show that external unlabeled data enables Unseen Knowledge Transfer (UKT), allowing the student to acquire knowledge from domains absent in its training data. Sequential distillation also causes Unseen Knowledge Forgetting (UKF), where transferred knowledge is lost after training on later teachers. To trade off between UKT and UKF, they propose Self External Data Distillation (SE2D), which…

MP-ISMoE framework uses Gaussian Noise Perturbed Iterative Quantization (GNP-IQ) to reduce memory and Interactive Side Mixture-of-Experts (ISMoE) to scale side networks, outperforming state-of-the-art METL methods on vision-language and language-only tasks.

This paper proposes SPARK, which addresses functional entanglement in LLM-assisted NAS by explicitly selecting functional factors for conditioned edits, achieving 28.1x sample-efficient speedup and 22.9% relative OOD accuracy improvement on CLRS-DFS.

This paper proposes an unsupervised method for categorizing transformations between pairs of inputs using parameter division. By splitting a single transformation's parameter into components and imposing homomorphism constraints, the method identifies normal subgroups without auxiliary assumptions. Evaluations on rotation, translation, and scale image pairs show that group-decomposition constraints drive appropriate categorization.

MetaAdamW is a novel optimizer that uses a self-attention mechanism to dynamically adjust learning rates and weight decay per parameter group, outperforming AdamW on five tasks with up to 17.11% training time reduction or 11.08% performance improvement.

This paper introduces a classification distinguishing scalar-reducible and scalar-irreducible dynamics, showing that scalar-irreducible dynamics enable internally generated regime switching through feedback between fast and slow variables, offering a path toward autonomous learning systems.

A new AI framework called ADAPTS uses a mixture-of-agents LLM architecture to automatically rate depression and anxiety severity from clinical interviews. It decomposes interviews into symptom-specific tasks, achieving expert-level accuracy and demonstrating potential for scalable psychiatric assessment.

A position paper argues that current AI systems should not be used for peer review due to a hivemind effect and vulnerability to gameable stylistic changes, calling for a dedicated science of peer review automation.

Terminus-4B, a fine-tuned Qwen3-4B model trained via SFT and RL with rubric-based LLM-as-judge reward, matches or exceeds frontier models like Claude Sonnet/Opus and GPT-5.3-Codex on agentic terminal execution tasks while reducing main agent token usage by ~30%.

A novel algorithm that learns correct sequential behavior from as few as 2-10 successful execution traces, combining dominator analysis with multimodal LLM semantic understanding to validate new executions. It achieves high accuracy with only 3 training traces across domains like UI testing, code generation, and robotic processes.

Proposes a framework to identify four types of mental model discrepancies in team dialogues: unsupported beliefs, false beliefs, belief contradictions, and omissions. Experiments show predictive signals in these patterns.

This paper proposes a novel LLM-based evolutionary search framework with programmatic context augmentation, enabling code-based interactions with datasets to extract rich information beyond scalar metrics, improving efficiency and accuracy in symbolic regression.

A new study reveals why many AI adoption efforts fail: workers, who are expected to collaborate with AI, are often overlooked in design and implementation decisions. Through interviews in healthcare, finance, and management, the paper identifies key barriers and proposes strategies to align AI with real-world practices.

This paper resolves the open question of computing Thiele rules on the voter interval domain by showing that the standard linear program admits an optimal integer solution, and provides a fast algorithm. It extends results to VCI and LC domains, proves LC strictly contains VCI, and shows that a tree-based generalization leads to NP-hardness.

This paper presents a tool-mediated LLM architecture that uses deterministic tools (Stackelberg best-response, Bayesian observer updates, attack-graph primitives) and finite action catalogs to provide formal guarantees for autonomous cyber defense. A composite Lyapunov function machine-checked in Lean 4 certifies controllability, observability, and Input-to-State Stability (ISS) robustness. Experiments on 282 real enterprise attack graphs validate the claims, and a tool-mediated Claude Sonnet 4 controller reduces attacker expected payoff by 59%.

CreativityBench is a new benchmark for evaluating large language models' creative problem-solving ability through affordance-based tool use. It includes a large-scale affordance knowledge base (4K entities, 150K+ annotations) and 14K tasks. Evaluation of 10 state-of-the-art LLMs reveals that models can select plausible objects but fail to identify correct parts, affordances, and physical mechanisms, leading to significant performance drops. Model scaling saturates quickly, strong reasoning does not translate to creative affordance discovery, and Chain-of-Thought yields limited gains. Creative tool use remains a major challenge.

Build American AI, a nonprofit linked to a super PAC funded by OpenAI executives and Andreessen Horowitz, is paying influencers up to $5,000 per TikTok video to spread messaging that frames China's AI advancement as a threat to American safety and jobs.

Image2Obsidian bridges the gap between physical notes and digital Obsidian vaults. It automatically detects AirDropped iPhone photos, performs OCR and semantic analysis using Claude's vision model, and creates structured Markdown documents in your Obsidian vault with the original image attached as a wikilink. Works via CLI or Claude Code skill, MIT licensed.

This post details how to build an automated company due diligence agent using LangChain's Deep Agents for orchestration and Parallel's Task API for structured web research, with five research subagents and compliance observability via LangSmith.

Budgetbreeze is an AI-assisted personal finance app that requires JavaScript to run.

Amp team completely rebuilds its CLI tool, launching Neo with remote control, automatic context compaction, message queuing and steering, new Plugin API, and major performance improvements.

A deep dive into the inference pipeline of Large Language Models, from tokenization and embeddings to self-attention and next-token prediction, illustrated with JavaScript code examples. Covers training (pre-training and alignment), scaling techniques (MoE, multimodal, quantization), and high-level applications like function calling, chain of thought, RAG, and agents.