The Archive

Theoretical analysis of convergence rates for asynchronous Q-learning under fixed behavior policies.

Constrained decoding with structural automata ensures well-formed semantic parse generation in DRS and AMR parsing.

Language model attention heads perform task-dependent roles via contextual network interactions, measured via influence and importance.

RATED framework combines semantic matching with contextual bandits to improve tool retrieval reliability for AI agents.

X-AgentRed red-teaming method discovers compositional failures in multi-service agent ecosystems via cross-service path exploration.

Theoretical characterization of memory state dimensionality requirements for continued learning in recurrent systems.

PISO optimization algorithm combines partial gradients with zeroth-order estimates for bilevel and performative prediction problems.

Stratechery weekly digest covering Apple, Facebook, and US-China dynamics; no AI-specific content indicated.

Finite-sample convergence bounds for discrete conditional diffusion models with analysis of score estimation and discretization error.

Inference-time steering with dynamic scaffold/epitope selection improves sample efficiency in de novo nanobody generation.

Amazon says it will stop using NDAs when negotiating data center deals with local governments, following a similar move from Microsoft earlier this year. Secrecy has fueled community backlash against AI infrastructure, with opposition leading to hundreds of proposed and enacted moratoriums from New York to San Francisco. Meanwhile, a wave of startups is betting that consumers will hand AI agents access […]

Go-Explore search over environment snapshots improves coding agent recovery; preserving runtime state (DBs, services) outperforms cold restarts.

PAGE framework applies vision-language models to large-scale comics and manga analysis, enabling panel/text parsing and historical research at scale.

For six years, Danu founder Amy Ma has been working on a better way to sort recyclable waste.

Models trained on chronologically ordered knowledge (specialists) update knowledge far more efficiently than models trained on shuffled data (generalists).

Systematic study of design levers in continuous diffusion language models, examining self-conditioning and embedding geometry for accelerated sampling.

"When we are drawn into even the most primitive exchanges with a relational artifact, we believe it cares for us," Dr. Sherry Turkle writes. "And we are wired to care for it in return."

SpatialASM framework enables LVLMs to internalize active spatial mentalization through training-time simulation, improving spatial reasoning via viewpoint reconstruction.

MC-EvoBench controlled benchmark audits embodied recursive self-improvement in agents, separating genuine capability acquisition from resource carryover.

OpenPA framework for free-form pronunciation assessment uses Whisper's cross-attention for joint word/utterance-level scoring without forced alignment.

Run-to-run variance in semantic query engines: 0.07–24% of rows change outputs across five LLM executions at temperature 0, with cascading amplification in plans.

Learned answer selector trained on Qwen3-14B improves abstention decisions for factual QA; coverage on TriviaQA increases from 20.2% to 30.9%.

SCCR audit framework quantifies predictive signal in raw-input models beyond expert features by selectively removing shared embedding directions.

TabMF repurposes pretrained tabular foundation models as dataset encoders via contrastive post-training for representation learning and distribution-shift detection.

Dream to Remember (D2R) addresses catastrophic forgetting in continual robot learning by steering frozen generalist policies with world models.

β-Flow unifies diffusion and flow-matching priors for linear inverse problems via exact Gaussian conditioning with correlation and confidence parameters.

QTestBench evaluates LLM-generated unit tests for quantum code beyond correctness, measuring ability to distinguish correct from faulty implementations.

SAFE-QE mitigates LLM hallucination in query expansion via subsequence-association filtering and breadth-aware selection without corpus grounding.

Applies combinatorial fusion analysis to diverse variational quantum circuits for improved predictions via rank-score and cognitive diversity scoring.

Phoneme-level error analysis of four data augmentation methods for low-resource ASR reveals complementary benefits beyond aggregate error rate metrics.