The Archive

Falsification benchmark reveals prediction bottlenecks do not recover causal structure; provides standardized test suite across architectures.

CIVeX verifies causal effects of tool-use actions in LLM agents via structural causal queries and identifiability checks.

WorldSpeech multilingual corpus: 65k hours aligned audio-transcript data across 76 languages for low-resource ASR improvement.

Looped-MoE transformers scale better than dense looped models; sparse layers enable routing diversity across repeated passes.

FORTIS benchmark evaluates over-privilege in LLM agent skills: minimal selection and boundary-respecting execution.

Full-prefix Matryoshka learning induces task-aligned privileged bases with per-dimension structure reflecting information content.

Persona vectors in LLM activations form dynamic polylogues during reasoning; polylogue features predict reasoning correctness.

Mixture policies in continuous RL offer theoretical flexibility but lack practical reparameterization tricks; study questions whether complexity outweighs benefits in state-of-the-art algorithms.

Deep learning framework analyzes grammatical gender shift from Latin to Occitan with improved tokenization for low-resource historical NLP.

Predictive model estimates large model pre-training loss from N, B, K parameters; outperforms Chinchilla on extrapolated compute budgets up to 1000x.

Hierarchical MARL architecture for traffic simulation combines multi-agent interaction reasoning with continuous trajectory planning beyond self-play equilibrium.

Meow-Omni 1: quad-modal MLLM for feline behavior analysis with high-frequency biological time-series data to decode animal intent beyond surface patterns.

AlphaExploitem extends AlphaHoldem for poker by learning to exploit suboptimal play beyond Nash equilibrium using hierarchical transformer reasoning over hand history.

User shares llama-server configuration for running Minimax 2.7 at 100k context on Strix Halo hardware with detailed tuning notes.

Empirical evaluation of LLMs vs. traditional taggers for POS tagging in Medieval Occitan, Catalan, French under zero-shot, few-shot, and transfer learning.

FedVSSAM identifies flatness incompatibility in federated SAM under data heterogeneity; proposes fix to align local and global flat minima in distributed learning.

[https://claude.ai/share/12659fcf-c1c8-4bbb-bc45-b41b26cd8b69](https://claude.ai/share/12659fcf-c1c8-4bbb-bc45-b41b26cd8b69)

Reddit discussion on safe autonomous agent architectures for personal assistants with tool access, exploring sandboxing, MCP, and approval-based patterns.

Federated learning evaluation for mammography under breast density heterogeneity; assesses robustness of FL algorithms in realistic multicenter clinical settings.

Reddit user seeks advice on LLaMA inference harnesses; discusses fragmentation and compatibility issues with local LLM tooling.

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.