The Archive

Data selection is critical for training large language models on massive and heterogeneous corpora. Meta-learning for Training-data Selection offers a principled alternative to heuristic scoring by learning data weights from a target validation objective, but existing methods face a trade-off between fine-grained valuation and transferability to unseen data. A natural solution is to replace per-sample weights with a selection network. However, we find that directly incorporating such a network into existing MTS objectives leads to unstable optimization and poor generalization, caused by weigh...

GeoLatent decomposes spatial latents for 3D reasoning in vision-language models using geometry-guided optimization to stabilize directional representations.

HumanoidToolBench introduces 18-task benchmark for tool use on humanoids with 3.1k real-world demonstrations from Unitree G1.

Kolmogorov-Arnold Networks applied to free-boundary PDEs via physics-informed residual loss functions, compared against PINNs.

Study argues Wasserstein gradient flows and forward-only diffusion cannot efficiently sample multimodal distributions despite theoretical convergence claims.

LLM2Jev framework extracts calibrated categorical decisions from LLM next-token probabilities without fine-tuning, enabling direct software integration.

Homomorphic Advantage Operator stabilizes RL under fully homomorphic encryption by mitigating Bellman drift from polynomial approximations.

PyPottery open-source suite automates ceramic documentation via image extraction, vectorization, and layout generation for archaeology.

A federal judge has dismissed a pair of antitrust lawsuits filed by Chegg and Rolling Stone parent company Penske Media Corporation, which accused Google of driving away web traffic with its AI-powered search features, as reported earlier by Reuters. US District Judge Amit Mehta takes Google's side in a ruling on Wednesday, writing that PMC and Chegg's claims don't stand up to antitrust law. In lawsuits filed last year, Chegg and PMC accused Google of abusing its monopoly power by coercing publishers into supplying content for AI Overviews for free, or risk disappearing from search results. T...

Causal Memory Policy restores identifiability of memory utility in LLMs by intervening on retrieval rather than storage decisions.

Probabilistic deep learning emulator models Holton-Mass stratospheric variability with interpretable latent structure for rare regime transitions.

Simon Willison releases pwasm 0.2a0, a pure Python WebAssembly engine improved via Claude Opus 5.5 code generation.

Theoretical analysis of sequential capacity bounds for quantum processes with fixed memory, establishing K log K growth law.

LoRA-Norm post-training method rebalances singular value gains in LoRA adaptations to preserve multi-task performance.

Span-level hallucination detection framework using hidden state probing to identify structured semantic drift in LLMs.

SimpleTimeBench diagnostic suite reveals zero-shot failures in Chronos-2, Moirai, Toto on basic temporal logic tasks.

OpenAI argues advanced AI's primary economic value lies in automating routine work supporting breakthrough ideas, not replacing innovation itself.

HydroJEV training-free classifier achieves one-second triage of SCADA alarms in water networks vs. LLM baselines.

Distributionally Robust Schrödinger Bridge learns single controller robust to shifts in initial distribution.

Study of cross-serial dependencies and stack-based LM learning biases on artificial languages with typological properties.

CARM masking scheme addresses off-policy response cancellation in LLM RL for reasoning and code tasks.

Mimir physics-grounded LLM agent learns long-horizon irrigation control with compounding errors and no reset.

Global Coherence Theorem bounds multi-agent coordination failure when indistinguishable observations require conflicting actions.

Transfer-and-ground framework generates human mobility trajectories in target cities using only POI data, no target trajectories.

Amazon Web Services' Strand Labs has released the latest Jevalike decision model, Strands Decider 2B.

Shopify’s new Canvas site builder lets merchants create and customize their online stores by chatting with its AI agent Sidekick, while watching the changes happen in real time.

Adding in a second neural network that guesses the identity of hidden pieces was key.

Albertsons Cos. deploys ChatGPT Enterprise and OpenAI API for internal team productivity and retail customer experience improvements.

Barclays expands Claude deployment across operations for client-facing services and internal workflows.

Brian Chesky on making Airbnb agent-friendly, the state of consumer AI, and why the world needs an AI-native operating system.