The Archive

OmniNFT applies multi-objective reinforcement learning to joint audio-video generation, addressing modality alignment and cross-modal synchronization challenges.

Needle: 26M parameter tool-calling model distilled from Gemini, runs 6000 tok/s prefill on consumer hardware.

MEME benchmark evaluates LLM agent memory across multi-entity and evolving dimensions, revealing system failures on dependency reasoning and deletion tasks.

Study reveals geometric coupling between routers and experts in sparse MoE models, explaining routing collapse and informing load-balancing improvements.

Framework identifies verifier failure and rubric design limitations as sources of reward hacking in RL post-training, tested against frontier evaluator panels.

KV-Fold enables long-context inference via training-free KV-cache recurrence, treating cache as functional fold accumulator over sequence chunks.

Attractor Models stabilize recurrent Transformers via fixed-point refinement with implicit differentiation, maintaining constant training memory across variable depths.

Theoretical work extends sample compression schemes to high-arity product spaces and proves connection to PAC learnability.

ScaleSearch optimizes Block Floating Point quantization scale factors via fine-grained search to reduce inference error for generative models.

Gymnasium environment for demand-response RL using offline smart meter data to optimize grid flexibility and energy affordability.

Proximal gradient sampler for composite log-concave distributions with convergence bounds in total variation distance.

Multi-Stream LLMs decouple single-message bottleneck into parallel streams for thoughts, inputs, and outputs, enabling concurrent agent reasoning and tool use.

llm 0.32a2 adds support for OpenAI's /v1/responses endpoint, enabling interleaved reasoning visibility for GPT-5 class models.

TextSeal watermark for LLM provenance and distillation protection uses Gumbel-max with dual-key generation and multi-region localization, zero inference overhead.

Real-world 5G/6G dataset for AI-driven beam management and handover optimization in mobile networks.

Computational study of LLM-generated political text detection across crisis events using behavioral analysis vs. perplexity signals.

GPT-5.5 outperforms Claude Opus 4.7 on ProgramBench coding benchmark, achieving first solve with fewer agent steps via action bundling.

ORCE method decouples verbalized confidence from answer generation in LLMs to improve uncertainty calibration without degrading accuracy.

The company named Open Doors Partners, Unicorns Exchange, Pachamama Capital, Lionheart Ventures, Hiive, Forge Global, Sydecar and Upmarket as companies that are not authorized to provide access to buy or sell its shares.

OpenAI CEO Sam Altman says Elon Musk did "huge damage" to the culture of the AI startup. During testimony as part of Musk's lawsuit against OpenAI, Altman said Musk required OpenAI president Greg Brockman and former chief scientist Ilya Sutskever to rank researchers by their accomplishments and "take a chainsaw through a bunch." Altman conceded that this was the management style the Tesla CEO was known for, but that it was incompatible with his startup. "I don't think Mr. Musk understood how to run a good research lab," Altman testified when his lawyer, William Savitt, asked about the impact ...

CLM detour improves domain-adapted encoder pretraining on biomedical texts vs. standard MLM continuation by 0.3-2.8pp.

CAAFC framework for automated fact-checking and hallucination detection aligns LLM-based AFC with professional fact-checker workflows.

Environment-adaptive preference optimization for rare-event prediction using long-tailed learning on wildfire datasets.

Google and SpaceX are in talks to build data centers in orbit, pitching space as the future home for AI compute, even as costs today remain far higher than on the ground.

(DISCLAIMER: I accidentally deleted the last post on this subreddit my apologies if this is your second time seeing it) Last year I made a [post](https://www.reddit.com/r/datascience/comments/1lkjxmr/steam_recommender_using_vectors_student_project/) about my steam recommender The last one was great and served its purpose of showing many people new games, But this new version is much more functional! I love making recommendation systems that tell the user WHY they got the recommendation. During a steam sale event, I always find myself trying to look for new video games to play. If I wanted ...

Reddit user spots "Claude Haiku 4.6" label on Anthropic tutorials page; likely a documentation error, now corrected.

Reddit user releases free mobile app for generating LLM wrapper applications locally.

Hugging Face releases physics-intern, a multi-agent framework for theoretical physics research that doubles Gemini performance on CritPt benchmark.

Reinforcement learning approach to construct minimally rigid graphs with high realization counts via Henneberg moves.

Theoretical analysis of geometric memorization in transformers showing embeddings encode relational structure vs. linear parameter scaling.