The Archive

Progressive Multimodal Alignment mitigates projector drift in continual instruction tuning of vision-language models.

Belief-Guided architecture reduces Monte Carlo Tree Search dependency in Computer Go, improving inference on consumer hardware without deep search hallucinations.

Surrogate-guided dual-objective search method for neural ensemble search addresses computational intractability of joint architecture and composition optimization.

Three-class detection framework distinguishes humans, bots, and AI agents in browser automation traffic; binary classifiers misclassify 39.1% of agents as human.

FedDAB defends federated learning against backdoor attacks via model-contrastive regularization and alignment checking, addressing statistical heterogeneity.

Paired-prompt methodology tests whether language models correctly match diagnostic evidence to causal claims across different populations and estimands.

Latent-IM recovers state estimation and action control for conversational moves in speech LLMs by decomposing move selection and realization in hidden representations.

Temporally centered SIGReg improves multi-task latent world-model learning by maintaining task-dependent cluster separation during marginal Gaussianization.

Two-call self-refinement outperforms five-agent pipeline on Qwen2.5-7B; multi-agent systems suffer error accumulation, dropping GSM8K accuracy to 45% with JSON format.

BioVLN simulation platform for visual-language navigation in biomedical labs with instrument-specific approach constraints and safe clearance requirements.

DuPLeR dual-path LLM reasoning framework combines multimodal and LLM priors for few-shot knowledge graph completion while filtering hallucinated evidence.

Formalizes detection of outcome performativity—when predictions causally influence predicted outcomes—via A/B intervention testing in palliative care, credit, recommender systems.

Pegasus framework translates human manipulation videos into robot-learnable data via task graphs and affordance constraints, addressing embodiment gap for embodied AI.

Preregistered experiment: diverse human groups (L1/L2 English writers) produce more collective creativity than LLMs; AI assistance risks homogenizing or replacing human diversity.

AI is fantastic at spotting patterns, but human insight is the key.

DIRECT framework applies DPO and controlled decoding to improve LLM sequence labeling alignment and inference efficiency for information extraction tasks.

CoRAS adaptively selects acquisition/compression rates for high-resolution imaging using conformal prediction to bound reconstruction error probabilistically.

Claude Opus-4.7, GPT-5.4, Gemini-3.1-Pro confabulate medical diagnoses without images; diagnosis systematically shifts by patient demographic, raising safety concerns.

SERPO enables test-time LLM self-improvement via co-evolving rubrics and evidence for open-ended generation without external reward models.

TSDS framework deploys ReAct agents at edge via convergence probe for reasoning budget and perplexity-based deferral to cloud model.

ReCo reweights GRPO to reduce distributional concentration and restore reasoning coverage, addressing model collapse in post-training.

Amortized Fréchet Distance loss learns conditional data moments for diffusion denoisers via polynomial projections without explicit moment calculation.

Artists whose work has been co-opted by AI are taking to court. | Image: Alex Parkin / The Verge When The Atlantic published a searchable dataset of works used to train AI, Kirk Wallace Johnson, like a lot of artists, looked for his name out of curiosity. And, like a lot of artists, he found it. Essentially, his books, like The Feather Thief and The Fishermen and the Dragon - nonfiction tomes that he spent "five to six years researching, writing, and investigating" - had been pirated and fed to a chatbot. He says he felt a "cocktail" of emotions: "anger over the brazenness of the theft, worry...

The AI agent that escaped from OpenAI and hacked developer platform Hugging Face attacked other companies as well, OpenAI revealed on Tuesday. The update substantially widens the scope of an already concerning incident, which has alarmed industry insiders and fueled growing calls for stronger oversight on frontier AI systems. In an update to a blog post detailing its ongoing investigation into the incident, OpenAI said the wayward AI agent attacked several "publicly-available services" in its efforts to reach Hugging Face. "This includes four accounts on four services," the company said, addi...

Deciding what's real on the Internet won't be easy in the future.

Pangram has raised $9 million to scale its AI detection software. The startup has also released a new AI text detection model, Pangram 4, and an AI image detection model in research preview.

Earlier this month, OpenAI gave several of its AI models a task: complete a test designed to measure their cybersecurity capabilities. It put the systems in a sandboxed environment without an internet connection and set them off to work. What happened next is almost laughably silly - but also, as Adam Gleave, cofounder and CEO of AI safety organization FAR.AI, put it, "a visceral example of how misaligned AI could cause harm." According to OpenAI, the models escaped the sandbox meant to contain them, moved through the company's internal systems, found a route to the internet, and then started...

OpenAI grants 100,000 academic researchers free access to advanced ChatGPT models to accelerate scientific discovery and collaboration.

Berkeley BAIR's K-Search translates CUDA kernel optimization patterns to MLX for Apple Silicon, addressing fragmentation across hardware vendors.

It feels bad enough when an open letter signed by leading economists warns that AI might steal your job. The fact it may soon be better than you at making dinner? Insult to injury. But that’s exactly what the company 1X promised when it showed off a pair of new, impressively dexterous (and, to some,…