The Archive

ReCIRC: input-conditional conformal risk control for segmentation/classification, improving on uniform-threshold marginal guarantees.

Framework uses MoE routing signals in Qwen3-VL to detect target-absence grounding failures in VLMs before generation.

Study explains how local mixing layers encode positional information in NoPE transformer attention without explicit position encodings.

Planning-as-Routing framework distinguishes selection vs execution failures in LLM agents across four planning modes.

iGSM benchmark reveals chain-of-thought traces often reach correct answers via invalid reasoning paths, questioning mechanistic interpretability claims.

Study shows audio encoder pruning in Speech-LLMs increases demographic disparities in WER across groups on Fair-Speech and Common Voice.

OpenAI's newly announced suite of office features puts it into mor direct competition with more traditional software companies.

Power-sharpened sampling balances exploration-exploitation at inference time to enhance reasoning in LLMs without RL post-training.

RICE enables training-free dense retrieval by prompting LLMs with in-context examples to produce effective query-document representations.

NeuronEye plug-in constructs sparse concept-level neuron vocabulary from VLMs to isolate and modulate query-relevant visual evidence.

Tail-Influence Sampling optimizes evaluation budget allocation for estimating CVaR in stochastic policies via tail-risk aggregation.

Probe-Space Preconditioning closes convergence gap between zero-order and backprop training, reducing OPT-30B memory from 600GB to 60GB.

Data-driven method for constructing surrogate models preserving dimensional homogeneity using Buckingham Π-groups and harmonic expansions.

Training risk-averse agents via character training and CARA-based constitution to mitigate misalignment harms through preference for safer strategies.

Mira system optimizes MoE inference on single-GPU systems using predictive expert staging and adaptive caching to reduce memory overhead.

Neural topology optimization for ship structures using Kolmogorov-Arnold networks to minimize vibration-induced compliance.

Hessian Null Space Continuation method maps mode-connected solutions in neural network weight space to study internal representation degeneracy.

"The chance of human extinction is about a coin flip, in my view," Geoffrey Irving, a former OpenAI and Google DeepMind employee, said in a new interview. It's one of a dozen interviews with AI researchers, including current and former employees at OpenAI, Google, and Anthropic. Palisade Research, which says it's a nonprofit studying AIs' capabilities and motivations, collected and launched the interviews on frominside.ai. Several of the researchers joined Irving in warning about the possibility AI could drive humans extinct. Neel Nanda, a Google DeepMind research scientist, said there's "at ...

Quantum learning theory: refutes classical conjecture on quantum-classical separations for exact learning with membership queries.

Method for quantifying LLM reasoning uncertainty using divergent token analysis beyond raw token probabilities for confidence calibration.

GEODE: foundation model for energy/radiation systems preserving heterogeneous scientific interfaces and native data representations.

Alpha Diffusion Language Models use sequence-level alpha loss to enable consistent parallel token generation in discrete diffusion.

Jaxolotl: high-performance benchmark suite for multi-task RL with LTL-based instruction following, standardizing evaluation and reducing compute.

Latent guidance method improves time series foundation model forecasts by adaptively ensemble-weighting in-context predictions.

Kernel optimal transport improves functional flow matching for generative models on infinite-dimensional function spaces.

Wabi is repositioning its prompt-based app builder as a personal AI agent that can create interfaces on demand, combining chat, apps and ongoing tasks.

UserProxyBench benchmark evaluates LLM-based user simulators in agent interactions via fidelity scoring independent of agent success.

Dots are meant to operate independent of any specific hardware or interface, pursuing user-defined goals continuously in the background with minimal oversight.

OpenAI is expanding ChatGPT plugins with dedicated sidebar homes, interactive panels, file viewers, improved discovery, and support for automations.

OpenAI says GPT-6.1 Sol delivers significant improvements over GPT-6 Sol across complex professional tasks, including code writing and debugging, document understanding, and executing multi-step business workflows.