The Archive

Theoretical paper derives unified KL identity for exponential families applicable to softmax, Gaussians, variational inference, and RLHF.

Linguistic study on syntactic dependency distance minimization in star-like sentence structures; narrow theoretical interest.

Differential privacy optimization for mean estimation in shuffle model; foundational theory without AI systems application.

DriftBench evaluates constraint adherence across 7 LLM models in iterative ideation; shows models lose fidelity under refinement pressure.

MIFair framework for bias assessment via mutual information; addresses intersectionality and multiclass fairness in ML systems.

TeCoD system improves Text-to-SQL accuracy via template-constrained decoding from query pattern reuse in labeled workloads.

https://preview.redd.it/u1ik0uejlcyg1.png?width=1080&format=png&auto=webp&s=d2ea7758fbfe5fdf2b65a3a79f2bb99711a07db8 As you can see in the outputs, Mythos can output images.

FedHarmony addresses label correlation drift in federated multi-label learning across heterogeneous client datasets.

Empirical study finds statistical laws in global recipe structures via NER; cultural/linguistic interest, not AI-relevant.

Cost-Aware SGD algorithm for finite-sum objectives with heterogeneous sampling costs; applied to RL with language models.

LLM-based combinatorial optimization for Design Structure Matrix modularization; engineering application without novel AI contribution.

Structure-aware densification improves 3D Gaussian Splatting convergence by distinguishing geometric vs. aliasing errors.

Anthropic just released Opus 4.7 as their most advanced model. I reverted to 4.6 within days. I use Claude for production work -- not chat, not summaries. Real deliverables with real deadlines. Here is what happened. I asked 4.7 to update a Word document. It is a task the previous model handled routinely. The new model produced a plain text markdown file with a .docx extension. Not a degraded document. Not a partially formatted document. A file that was literally not a Word document at all. Delivered with full confidence and zero warning that anything was wrong. When I caught it and ...

Framework reframes clinician AI overrides as preference learning signals for RLHF in clinical decision support, with observable outcomes and expert annotators.

Kernelized advantage estimation reduces RL training overhead for LLM reasoning by replacing value networks with nonparametric statistics.

Architectural pattern language for vision language agents balances latency/non-determinism of VLMs against real-time enterprise control requirements.

Latent-GRPO stabilizes reinforcement learning in latent reasoning space by addressing probability density and sampling mechanism shifts.

Comparative evaluation of three LLM agent paradigms (domain-specific, computer-use, coding) on scientific visualization tasks across 15 benchmarks.

what could you possibly be thinking so long for 😭

Dynamic scaled gradient descent prevents optimization collapse during fine-tuning on imbalanced datasets via gradient scaling.

X is rolling out a rebuilt ads platform powered by AI as it works to grow revenue again.

ITS-Mina applies Harris Hawks optimization and external attention in MLP-based framework for multivariate time series forecasting.

D3-Gym dataset: 565 verifiable tasks from real scientific repositories for evaluating LLM agents on data-driven discovery.

TransVLM formalizes shot transition detection (not binary cut detection) for video editing using vision-language models.

LLM and diffusion models generate procedural trading card game content to improve metagame diversity in Pokémon.

MLLMs fail on circuit-to-Verilog translation due to 'Mirage' phenomenon; visual perturbations cause hallucinated code despite correct diagram interpretation.

TMLR paper introduces Joint Embedding Variational Bayes, a probabilistic framework for non-contrastive representation learning via factorized embedding likelihood.

Despite only having one face, I made testing work. I'm currently wearing a pair of smart glasses called the Even Realities G2. Another two pairs, from Rokid, sit on my desk. A few feet away, I've got the Meta Ray-Ban Display charging alongside their Neural Wristband. In my closet are six pairs of $50 smart sunnies that an overzealous Walmart rep sent me. Those sit next to some Xreal, RayNeo, and Lucyd glasses, plus an old pair of Razer Anzu. Later, I'm calling my optician because I'm hoping to test a pair of the new Ray-Ban Meta Optics, which can supposedly handle my challenging prescription....

StructGP uses differentiable DAG learning on irregular EHR time series for uncertainty-aware clinical forecasting with interpretable causal structure.

Reddit user reports DeepSeek model refusing to acknowledge Taiwan's sovereignty, raises questions about geopolitical bias in AI systems.