The Archive

Mia Goth at The 32nd Annual Actor Awards. | Variety via Getty Images Earlier this month, news broke that actress Mia Goth had filed for divorce from actor and filmmaker Shia LaBeouf. An Instagram post from the account miagoth_ - marked with a blue verified checkmark - quickly spread. The post shows Goth in a T-shirt that reads "MY EX IS MY BIGGEST FAN" in bold black letters. The Beyoncé song "Single Ladies (Put a Ring on It)" had been added to the Instagram post. Hundreds of thousands of people - including some in the entertainment industry - liked the image, and entertainment outlets like OK...

Varest compresses image tokens via semantic variable-resolution tokenization and learned relevance pruning, maintaining visual quality across three backbones.

RadarGSOcc predicts 3D semantic occupancy for autonomous driving using 4D radar tensors with Doppler-aware geometric-semantic feature extraction.

Grounding Awareness enables personal agents to distinguish accepted vs. unacknowledged user arrangements, addressing memory-action alignment in multi-turn interactions.

ProAS resolves credit assignment in LLM agent RL via progress-aware advantage shaping, improving over uniform critic-free methods without adding critic overhead.

Generative Trajectory Planning balances inference efficiency and long-horizon reasoning by sampling action sequences from learned proposal distributions.

Contextual entrainment in LLMs involves shared interventions across relations, not isolated relation-specific heads, affecting robustness to irrelevant context.

SMI constructs ontology-grounded semantic IDs for biomedical LMs, improving hierarchy-aware representation of rare clinical codes without retraining.

Google DeepMind releases Gemini 4 Argon optimized for robotic control and physical reasoning tasks.

Gemini 4 Argon announced as high-capability model targeting complex problem-solving across domains.

Developers will get 50 percent discounted API rates through November.

After months sitting on the sidelines of the frontier AI race, Google has finally launched its long-awaited Gemini 4 Argon model, claiming it can compete with the best the industry has to offer. Google says Argon delivers "frontier reasoning and improved performance across complex workflows," while cutting down on hedging, repetition, and the cliches that continue to plague AI chatbots. The company claims Argon outperforms rival models like GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 across a range of benchmarks, including coding, cybersecurity, science, finance, and law. Google says i...

We’re giving developers, enterprises, scientists, and consumers access to Gemini 4 Argon, our latest frontier model.

Solar geoengineering simulation shows what happens when geopolitics takes over.

Multilingual dense retrievers exhibit systematic retrieval divergence across languages in shared index space, caused by per-language ranking conflicts, not weak understanding.

Epic Games plans to use AI to generate code and art for an unannounced Disney game, which is now delayed from its original March release window, according to a report from Bloomberg. The project's developers have been given an "AI-first" mandate directing them to use tools like Anthropic's Claude more to finish the game. According to Bloomberg, one of the game's directors said in a message to staff, "No more hand-written code." Staff were also told that all of the game's art will be "generated with full AI workflows." In a statement given to Bloomberg, Epic Games spokesperson Jake Jones said ...

APTree framework automates robotic task planning in construction via behavior trees and modular symbolic planners with runtime replanning.

LLM agents fail to load relevant skills at scale: implicit routing over skill metadata causes skill-load recall to collapse from 73.6% to 27.4% with 1,000 irrelevant skills.

Language model logits follow consistent log-linear geometry across contexts and scales, emerging during pretraining and constraining decoding behavior.

Knowledge distillation imprints teacher behavioral tendencies (harm-acceptance/avoidance) into students beyond task capability on unrelated reasoning tasks.

CACTUS pruning method preserves influence functions and gradient alignment in compressed models via curvature-aware objectives.

Study across 5 LLMs and 193 countries shows models approximate country-level time preferences correlating with GDP_PPP and income, suggesting embedded cultural assumptions.

Asynchronous multi-player multi-armed bandits with heterogeneous activation probabilities: theoretical analysis of optimal decentralized allocation.

Building optimization benchmark and foundation models for HVAC efficiency targeting 40-60% of building emissions; identifies policy transfer as key obstacle.

Computer Use Agents lack information-flow security against indirect prompt injection due to visual interface ambiguity; proposes structured threat model.

Empirical study on coding agents' ability to digitize clinical trial protocols from unstructured PDFs into structured machine-readable data.

LlMBO applies Bayesian optimization to fine-tune LLMs with minimal parameters, improving on TinyLoRA's gradient-based approach in low-dimensional settings.

EgoOmniWorld benchmark unifies egocentric video understanding via ego-action-world interaction loops across modalities and evaluation formats.

Study on faithful autoformalization at theory scale, proposing automated evaluation methods that trace individual expressions and contextual interpretation.

Analysis of tool-call judges' derivation gap: LLM agents' proposed argument values are rejected more often when absent from reference text.