The Archive

ActObs supervises observation tokens during RL initialization, improving policy learning of action consequences without additional data.

Paper proposes 'summarization bias'—LLMs collapse narrative structure into abstract summaries rather than reconstructing inferential meaning—with formal testing protocol.

Theoretical optimization paper on first-order oracle complexity for Lipschitz convex functions over norms; resolves COLT 2015 open problem.

TetrisCNN model interprets phase transitions in quantum simulator data without a priori physical knowledge; improves robustness on noisy experimental data.

Theoretical optimization: first-order black-box convex optimization over ℓp-balls; matches lower bounds and improves Euclidean-Lipschitz rates.

HerHealthEval benchmark: multilingual, register-sensitive evaluation of LLM understanding in women's health across English, French, Modern Standard Arabic.

Survey: RISC-V open processor ISA in ML applications; analyzes ecosystem from instruction extensions through deployment, commercial and academic implementations.

Causal discovery in DAGs with mixed data types (ordinal, count, continuous) using ordered logit and exponential family models; addresses identifiability.

HIL-UMI: human-in-the-loop post-training for vision-language-action robot models on Universal Manipulation Interface via interactive policy refinement.

Qualitative study: people retain ownership of AI-assisted work when they lead, iterate, or rewrite; lose it when merely approving AI suggestions.

FL-Net: federated learning framework for multi-center clinical research with data harmonization, disclosure control, versioning; evaluated against 14 existing frameworks.

Framework unifying physics-informed neural networks and finite element methods via Gauss-Newton discretization for PDE solving.

Probabilistic extension of Allen's interval algebra for uncertain temporal relations in language and perception tasks.

Model maker commits to new framework for reporting misaligned models

UniPolicy: multi-objective alignment framework for search advertising balancing relevance, click propensity, and commercial value.

Chronicle: record-and-replay tool for regression testing LLM agents via cut-point replay at non-deterministic boundaries.

Simulation platform testing LLM-based diagnostic strategies for autonomous underwater vehicle failure recovery without comms.

Practical fingerprinting attacks exploit frontier models' ability to trigger inference engine vulnerabilities via crafted tokens.

AMPLE-Math benchmark isolates privileged information contribution in on-policy self-distillation via 5,319 math problems.

Anthropic launches Life Sciences Verification Program granting researchers access to Claude models with relaxed safeguards for biology applications.

COIN-GP: distributed cooperative learning with Gaussian processes for sensor networks with partial state observations.

Recursive quantum LSTM architecture compared to standard QLSTM for one-step temperature forecasting on Toronto weather data.

Study shows explicit sense inventories improve LLM performance on Word-in-Context lexical disambiguation vs. traditional WSD.

Multi-objective Bayesian optimization for crystal structure refinement using X-ray and neutron diffraction data.

SAFARI: industrial benchmark evaluating LLM reliability in automotive hazard analysis under ISO 26262 safety standards.

Theoretical analysis of discrete diffusion models showing per-position distributions cannot match dependent token distributions.

NS3Learn transfers 5G NR sidelink reception realism to traffic simulation for connected-vehicle safety testing.

Theoretical analysis of TAP approximation and posterior geometry in spherical linear models under Marchenko-Pastur conditions.

SGPS combines sampling-based MPC with first-order policy gradients to improve visual policy learning for robotics.

StratCBT dataset of 9,688 psychological counseling sessions with CBT strategies for training dynamic LLM counselors.