The Archive

Double-Stitch learns population dynamics in Wasserstein space without simulation, enabling efficient training for conservative and periodic processes.

FrugalEvo optimizes LLM-guided evolutionary search by coupling high-cost and low-cost models with cache-efficient iteration to maximize performance per unit cost.

Planning to Learn shows policy-gradient methods in classification underperform cross-entropy because exact gradients are myopic; proposes forward-looking objectives for post-training.

Pivot-SD improves masked diffusion language models via self-distillation by selecting high-impact token commitments during denoising for targeted training.

In a throwback to MySpace-style internet drama, Shivon Zilis announced that she and the father of four of her children, Elon Musk, had broken up on X. To do so, she quote-tweeted a post from "Big Tech Alert," an account that, among other things, monitors what accounts are following and unfollowing each other. That post noted Musk was no longer following her on X. It was also sponsored by Kalshi, because everything in this relationship is totally fucking cursed, dude. In her statement, Zilis wrote, "It's hard to go from in love to let go in a week with no warning, but that's just how it is som...

Method uses simulator rollouts to train prediction models for new policies, addressing cold-start by learning policy-dependent dynamics before real deployment.

Compares STFT and harmonic CQT representations for multi-pitch estimation in vocal ensembles, finding efficient STFT alternatives to expensive feature extraction.

MRVQ enables elastic dimension and rate control in dense retrieval via residual quantization, allowing single index to serve multiple latency/quality/memory tradeoffs.

PoCoFL adds verifiable compliance to federated learning, enabling policy-compliant client contributions and aggregator validation across diverse FL topologies.

Proposes onboard anomaly detection for Earth observation satellites using self-supervised encoders to compress hyperspectral imagery for marine hazard detection.

Improves Gaussian Process Latent Variable Models via amortized structured variational inference to better capture uncertainty in learned low-dimensional manifolds.

E-process-authorized Thompson sampling permits controlled forgetting in non-stationary bandits by authorizing discounted states before departing from optimistic Thompson sampling.

Theoretical analysis proving convergence of success conditioning to optimal policies in MDPs with derived convergence rates.

IDRF framework enables efficient reward fine-tuning of masked discrete diffusion models via inverse-distillation regularization.

Analysis of scale symmetry breaking in undercomplete linear autoencoders under finite-stepsize SGD.

Expert-validated benchmark of 1,042 items evaluating 15 LLM models on Colombian legal system knowledge.

LoGo framework uses spatially localized rewards to improve 3D consistency in long-horizon camera-controlled video generation.

Dependency-Aware Group Policy Optimization traces read-write dependencies for improved credit assignment in terminal-using agents.

World Embedding Benchmark with 8,000 physics-grounded simulation cases evaluates how video representations encode physical properties.

NeutronGym provides executable physics-graded environment for LLM agents to design neutron instruments via McStas simulation.

Empirical finding that denoising computation in multi-step diffusion unfolds across network depth in compressed one-step models.

FALCON framework generates realistic, ambiguity-aware NL-to-SQL synthetic training data matching real-world query complexity.

First efficient algorithm for computing normal-form correlated equilibria in finite-horizon Markov games with n players.

UniIntervene++ adaptively allocates control between autonomous and assisted behaviors during online robot reinforcement learning.

Emotion regression from EEG-fNIRS data during video viewing using low-cost video-time priors as baseline.

Wayfarer: general option discovery method for accelerating reinforcement learning in Atari without handcrafted representations.

PPO agent trained on analytically-solved broker-trader game to evaluate RL convergence against known optimal controls.

Path integral surrogate method for defending against multi-step gradient inversion attacks in federated learning.

Stratechery weekly roundup touching Meta strategy, OpenAI developments, and cultural commentary; lacks specific AI claims or announcements.

With all the talk about how AI might one day ill us all, it's easy to forget that AI has already harmed some people, psychologically. Circuit Breaker Labs has created "crash test dummies" to solve that.