The Archive

SpecMRL improves Matryoshka Representation Learning by reorganizing feature space with spectral basis ordered by importance for better compact encodings.

HoloCosm benchmark compression method selects samples across multiple metrics, showing single-metric selection poorly preserves rankings on other dimensions.

Impossibility result shows ETF target geometry incompatible with analytic continual learning; proposes equinorm method for imbalanced settings.

TinyLlama-1.1B's tokenizer severely penalizes Indic scripts (22.6× subword bloat for Kannada vs. 1.34× for Latin), undermining reported efficiency metrics.

Proposes Martingale Posterior Cox Processes for Bayesian nonparametric inference on event data, primarily theoretical contribution with limited LLM relevance.

Fisher-Geometric Pruning post-hoc reduces overparameterized symbolic regression models by exploiting information geometry.

Study reveals token count and aggregate accuracy mask true VLM pruning quality; proposes per-example and latency-aware evaluation metrics.

ServiceBench benchmark evaluates LLM agents on complaint handling across 2,500 scenarios; finds current agents struggle with emotional reasoning under information asymmetry.

SO(3)-equivariant CNN on sphere for physics datasets respects rotational symmetry, reducing simulation requirements.

Analysis of primal-dual optimization for constrained MDPs under two-timescale non-stationarity, modeling oscillating constraint levels and sensitivity.

Gender inference model for Chinese names in bibliographic records addresses scientometrics bias but lacks AI architecture novelty.

Neural Expansion framework explains deep network generalization via ray coverage geometry, unifying margin bounds and neural tangent kernels mechanistically.

Blitz distillation reduces AlphaFold3 and Boltz-2 denoising from 200 to 8 steps via geometric noise schedule and multistep consistency training.

PresenceBench benchmark measures LLM adaptation to user functional state (readiness, flow, stability) via receptivity and specificity metrics.

Theoretical analysis of ranking tradeoffs among classification metrics in binary classification tasks.

Non-parametric stochastic optimization framework using Malliavin calculus for time-varying decision problems.

METR will publish its own assessments of Anthropic’s risks, incidents, and practices, and the frontier AI model maker has given up the right to block them.

GeoState decouples geometry features from camera viewpoint in 3D world models, separating scene dynamics from view-dependent variation for improved predictability.

ACEdit stabilizes multi-turn image editing by training on accumulated distribution shifts from model's own prior outputs.

AVC-Reward proposes event-centric evaluation framework for audio-video generation models handling multi-event composition and asymmetric cross-modal relations.

Topology-aware loss functions for wind-field super-resolution favor different model architectures than pointwise metrics.

Anthropic grants METR independent evaluator access to Claude systems; exploring partnerships with Hugging Face and others for safety assessment.

LOADBench: open dataset and standardized metrics for ML-based analog IC design, covering 100+ opamp topologies in SkyWater130 PDK.

Analyzes Muon optimizer optimality via Random-Radius Isotropic Curvature Model, introduces geometric conditions for polar-factor update stability.

SITAR is a situational distortion metric for ASR evaluation that captures semantic consequences of transcription errors beyond lexical differences.

Bayesian online bandit algorithm selects classifiers balancing predictive performance and fairness subgroup metrics with delayed label feedback.

semCER metric weights OCR character errors by severity via LLM judge, outperforming uniform Character Error Rate across English, Vietnamese, Chinese.

Nonparametric latent semantic model proves identifiability of three preference-disagreement types (directional preference, persistent disagreement, non-binary mass) from pairwise feedback.

Curvature-aware adaptive perturbation improves Sharpness-Aware Minimization (SAM) generalization by adjusting perturbation radius based on local loss landscape geometry.

Cortex consolidates context-specified knowledge into LLM parameters via prospective queries and distillation, unifying context editing, skill learning, and behavioral steering.