NLP & ML
Natural language processing, machine learning and deep learning.
15 published articles
Federated Learning
When users can't be shared, FedCGR shares a language instead
FedCGR treats federated cross-domain recommendation as generation over a stable semantic item language. Items become discrete semantic IDs drawn from public metadata, so clients align without exchanging private interactions. The trade-off: a semantic-only bottleneck that local collaborative filtering evidence must fill.
2026-08-18
Knowledge Graphs & Urban AI
Stations aren't islands: how RTSKG rewires urban transit data
City-scale transit models often ignore how stations relate to roads and businesses. RTSKG, a knowledge graph dataset posted to arXiv, models those interactions explicitly and reports gains on store recommendation and ridership prediction.
2026-08-18
RAG research
BM25 beats agentic RAG when corpora pass 10 million tokens
Agentic search wins on small corpora, but a muset-ai study across 28 nested tiers shows BM25 overtaking it near 10 million tokens. The agent burns 39x more query tokens, and graph RAG stalls in construction.
2026-08-07
AI Research
Voice Memory: a 776-byte file that tells speech recognition when to do nothing
A new inference-only scheme for speech recognition learns restraint: a frozen corrector reads a per-domain memory file and decides when to abstain. Unconstrained correction breaks correct tokens on up to 64% of edits; Voice Memory cuts that to 35% and lowers weighted WER from 8.36% to 7.52%.
2026-08-06
Research
Shared memory helps AI agents, but only until you hit five of them
Researchers shared replay buffers across actor-critic agents on parameterized action tasks. GAC jumped in performance, but SAC and TQC only crept forward. Beyond five agents, the computational cost climbs with no meaningful return. The paper offers a practical boundary for how far shared-experience methods can stretch before they stall.
2026-07-27
AI & Neuroscience
An AI model's scrambled neurons just recreated a stroke's exact damage to speech
Researchers perturbed LLaVA 1.6 to simulate aphasic picture-naming errors and matched individual patient profiles in up to seven error categories for 79.5% of cases. Six out of seven error types emerged naturally across different perturbation configurations, suggesting general-purpose multimodal models can function as digital twins for post-stroke language deficits.
2026-07-24
Embodied AI
Xiaomi just gave every robotics lab a 38-billion-parameter data factory
With a 38-billion-parameter model that unifies four robot tasks and open-sources the entire pipeline, Xiaomi aims to break the data bottleneck in embodied AI. Early benchmarks show a 26% improvement in task completion rates when training on model-augmented data.
2026-07-15
Physical AI
Alibaba's Qwen is now the brain inside 150,000 robots, cars, glasses, and drones
Alibaba's Qwen AI family now powers over 150,000 hardware devices, from humanoid robots to children's cameras, as the company pivots from chatbots to Physical AI, integrating multimodal models into robots, cars, glasses, and drones.
2026-07-14
AI Research
GPT-5.5's big win reveals something missing from every agent benchmark
EvoPolicyGym isolates a critical but understudied capability: an agent's ability to refine an executable policy through repeated feedback-constrained edits. The benchmark reveals GPT-5.5 as the strongest performer across 16 environments, and provides trajectory-level diagnostics that expose how different agents allocate budget and convert feedback into tuned parameters.
2026-07-11
Memory management
Bing found a 12,000-line allocator that breaks a 50-year-old trade-off
Microsoft Research's mimalloc allocator uses thousands of per-page free lists and a clever page-stealing technique to achieve both high concurrency and low memory overhead, breaking a trade-off that constrained systems for decades. A deep-dive from the RiSE group reveals the design.
2026-07-10
NLP history
The benchmark that made language models speak: how 2018's glue bet changed ai forever
The GLUE benchmark, launched in 2018, transformed natural language processing by providing a standardized yardstick for language understanding. Its legacy lives on in every modern LLM benchmark, from SuperGLUE to the latest arena-style evaluations that define today's AI race.
2026-07-08
Efficient World Models
Fast-LeWM just made visual planning stop stumbling over its own steps
Researchers introduce Fast-LeWM, a latent world model that speeds visual planning by predicting future states from action prefixes in parallel. The approach cuts computational costs and error buildup, outperforming prior one-step transition models.
2026-07-07