The Archive

Andreas Påhlsson-Notini argues current AI agents inherit human flaws—lack of rigor, patience, focus—rather than transcending them.

CoCo-SAM3 resolves mask overlap and semantic drift in open-vocabulary semantic segmentation by modeling concept conflicts across prompts.

Empirical study on prompt design vs. model selection for predicting fan experience ratings from survey text using GPT variants.

Google announces Deep Research and Deep Research Max features for enhanced information synthesis.

Plasma 1.0: 235M-param LLaMA-style model trained from scratch on single RTX 5080 GPU.

Introduces micro language models (8M-30M params) that generate first tokens on-device while cloud model completes response, masking latency.

Proposes Penalized Predictive Control framework using online Riemannian optimization for safety-critical control with black-box world models.

SafetyALFRED benchmark evaluates multimodal LLMs (Qwen, Gemma, Gemini) on hazard recognition and mitigation in embodied kitchen environments.

Introduces Chunk-wise Interleaved Splicing paradigm enabling autoregressive models for real-time streaming target speaker extraction.

Proposes TSAG framework and benchmark for evaluating LLM reasoning on financial time-series analysis with task delegation to computation modules.

You just cannot convince me this is how people order coffee. | Image: Starbucks Venti iced coffee, light skim milk. That's what I get at Starbucks. It is what I have gotten at Starbucks every time I've been to Starbucks for as long as I can remember, other than a brief love affair with the caffe misto a few years ago. In person, my brain barely needs to activate to say the words aloud; in the app, it's four taps and I'm ready to go. My first time ordering Starbucks through its new ChatGPT integration, which launched last week, was comparatively a complete mess. Getting started is easy enough,...

SAGE method for edge-cloud inference selects semantic content beyond attention importance under hard uplink bit-budget constraints.

Release of German free-association norms dataset for 5,877 cue words as part of multilingual Small World of Words project.

Benchmark comparison of Kimi K2.5 vs K2.6 on MineBench; K2.6 shows inconsistent but high-ceiling performance at $2.35 cost.

Kimi K2.6 vs K2.5 benchmark on MineBench shows cost-effective performance gains at $2.35 total.

Latitude's new AI-native platform, Voyage, aims to help gamers create their very own role-playing game.

User reports Image Gen 2 demonstrates advanced reasoning in color composition, generating cinematic palettes and tonal coherence across panels.

AblateCell agent reproduces baselines and systematically ablates components in AI virtual cell repositories for biological research.

Simon Willison comments on synthetic data injection into training sets via absurdist example (pelicans on bicycles).

Comparative study of output consistency across GPT-4.1, Claude Sonnet 4.6, Gemini 2.5 Flash for exercise prescriptions with safety analysis.

Romanian legal domain GEC dataset addresses grammatical error correction in low-resource language niche; narrow domain applicability.

Theoretical reduction from online learning to multicalibration via EVI solvers; advances in regret minimization analysis.

User documents that Gemma 4's default vision budget (280 tokens) is too low; Variable Image Resolution mode requires manual tuning.

Corpus of Sardinian extemporaneous poetry for NLP; minority language resource with limited frontier AI relevance.

Empirical study: LLMs measurably alter peer review opinions on clarity, originality at top AI conferences.

Framework for long-horizon terminal agents using observational context compression; reduces quadratic token cost via self-evolving compression.

Bond wants you to get off the couch and get back into the real world, its creator says. The new platform's AI system is designed to motivate users to do things away from the app.

Lyapunov stability analysis of constant-stepsize Q-learning via switched systems; theoretical convergence bounds.

RL post-training for visual semantic arithmetic in multimodal LLMs; extends relational reasoning from text to images.

Interpretability method for ColBERT retrieval via learned latent space reference; targets biomedical/clinical ranking diagnostics.