The Archive

BoostAPR applies RL with dual reward models (sequence-level and line-level) to automated program repair, enabling credit assignment to critical code edits.

MCP-Cosmos integrates world models into Model Context Protocol agents to bridge planning-execution gap via predictive task automation.

Data-driven circuit discovery tests whether LMs implement single computational subgraphs per task, challenging hypothesis-driven interpretability methods.

Controlled study shows external evolution outperforms internal deliberation (p<0.01) for multi-agent constitutional rules in coordination tasks.

Cosine Gated Adam Decay optimizer improves asynchronous DiLoCo training by scaling stale pseudo-gradients using exponential decay.

Diffusion models with MCMC accelerate low-thrust spacecraft trajectory design by learning high-quality initial costate distributions.

Apple discontinues 256GB M3 Ultra Mac Studio config; Reddit speculation about M5 Ultra memory trends.

Framework maps LLM reliability techniques (retry, voting, self-consistency) to Shannon coding theory operators as stochastic channel reliability methods.

DeepMind employee argues private AI labs should go public or allow retail investment to avoid enriching only billionaires.

Characterizes user-diversity conditions for O(1) regret and log(1/ε) sample complexity in personalized LLM alignment.

Empirical study comparing quantum vs. classical transfer learning robustness under reduced training data.

AI-native security framework structures asset prioritization across cloud/identity/config signals using exposure and business context.

Reddit user describes workflow challenges using Claude for bulk PDF document review and legal complaint triage.

Contextual Plackett-Luce model handles ambiguous sequence selection by matching multi-modal target distributions despite single-instance supervision.

Empirical comparison of expert-guided RL methods on continuous control with shared benchmarks, revealing three failure modes missed by isolated evaluations.

Reddit discussion about fictional robots people want in real life; mentions Disneyland R2D2 replica ($20k) and hypothetical T-800 chef.

Fin-Bias benchmark evaluates LLM decision-making in finance under human bias and uncertainty, addressing reliability concerns in financial deployment.

Community member built Autoharness, a Claude-powered tool that auto-optimizes agent harnesses (prompts, hyperparameters, scoring) via eval-driven iteration, achieving 40.7% improvement.

First comprehensive survey unifying token economics for LLM agents, framing tokens as production factors and analyzing computational-economic trade-offs.

User reports improved Claude iOS app performance enabling multi-agent workflows and sustained productivity on $100 plan without rate limiting.

GRC unifies generation, retrieval, and compression in single LLM forward pass using meta latent tokens for long-context agentic tasks.

Dynamic Meta-Metrics proposes source-sentence conditioned weighting for machine translation evaluation across language pairs.

SpectraNet combines spectral convolutions with U-Net hierarchy for stable autoregressive PDE surrogates, addressing rollout-error growth.

Extends bilevel optimization to multi-task setting with relaxed lower-level convexity assumptions for modern ML complexity.

Character-level Transformer for Tajik-to-Persian transliteration with 52K-word parallel corpus from verified lexicographic sources.

Studies approximation behavior in visual grounding under mismatched captions via controlled counterfactual perturbations to improve robustness.

FLiD: field-localized forgery detection framework for digital identity documents targeting critical regions rather than full-document processing.

Diagnosis framework identifies acoustic representation bias in audio deepfake detectors (AASIST, Wav2Vec2+ResNet18) beyond training data imbalance.

Constant-Target Energy Matching (CTEM) unifies density estimation across continuous, discrete, and mixed-variable domains via energy-based framework.