The Archive

Novel GNN-based framework for multivariate time series anomaly detection using prior-observation adversarial learning.

Labor data shows job displacement accelerating in customer service, admin, sales roles as businesses deploy AI solutions at scale.

Been writing backend for 11 years. last 8 months I've moved most of my work into claude code. I want to ask something and I'm not sure how to phrase it. when I spend a full day in claude code and ship 3 or 4 PRs, do I actually feel like I worked? or do I feel like I supervised? its not the same thing as a "did I solve hard problems today" question. its something weirder. I shipped real code. tests pass. PRs got merged. by every external metric the day was productive. but I cant point to a single moment where I thought hard about anything. I was just reading claude's diffs and going "yep...

Community speculation on open-weights model release cadence post-May/June 2024, questioning if release patterns are shifting.

iPOE system for interpretable prompt optimization via explanation-guided search to identify performance-driving changes.

BanglaMedVQA: first clinically validated medical visual QA benchmark for Bangla language with LVLM evaluation.

TaskGround framework for household agents to infer executable task structure from complete scene context and situated requests.

Symmetry-compatible optimizer design principle ensuring equivariance under parameter space symmetries for embeddings and MoE routers.

Analysis of spatial bias in German newspaper coverage of landslides, correlating media attention with geographic susceptibility.

Safety Geometry Collapse: multimodal LLMs fail safety transfer due to representation drift; proposes adaptive correction mechanism.

SENSE framework combining satellite imagery and generative models for urban building energy synthesis and functional layer generation.

Novel framework for routing problems with asymmetric hybrid geometries (points, lines, areas) using learned representation and decision-making.

User reports using Cowork browser extension to automate data removal from brokers; anecdotal testimonial with no technical depth or industry impact.

PARAM-Delta method upcycles dense models to MoE architecture for efficient multilingual LLM expansion via post-training without full alignment.

Python package for reduced-order modelling of multi-physics nuclear engineering problems using data-driven techniques.

Parameterized 4-qubit quantum game circuits for innovation recommender systems using Horizon Europe funding data.

Theoretical analysis of softmax transformers with low-precision quantization and Chain-of-Thought, proving Turing completeness with logarithmic scaling.

Analysis of equilibrium selection in multi-agent policy gradient methods via basin-entry probability and peer-learning decomposition.

LMAC system uses LLM reasoning to design communication protocols for multi-agent RL, improving state reconstruction under partial observability.

A-ProS agent solves competitive programming via multi-model feedback loop, separating solution generation from execution-driven refinement.

KVDrive multi-tier KV cache system spans GPU/DRAM/SSD for long-context LLM inference, addressing decoding latency via intelligent offloading.

Theoretical bounds on DDPM sampling error via Föllmer process, achieving optimal dimension and step scaling for Wasserstein distance.

Teger framework reduces error compounding in recurrent time-series models by modeling spatial-temporal residual correlations.

PPAI enables peer-to-peer collaboration between personalized LLM agents on edge devices via task delegation based on specialization.

MixCount dataset addresses mixed-object counting in real-world settings, combining real and synthetic data to improve model generalization.

FLAG applies diffusion models to spatial gene expression prediction by treating it as structured distribution modeling rather than pointwise prediction.

Study shows KV cache eviction policies require structural protection at prompt boundaries; 10% reserved cache recovers 69-90% quality on long-context models.

DocOS framework enables GUI agents to proactively retrieve procedural knowledge from documents to handle long-tailed automation tasks.

Study evaluates whether predictive uncertainty in robotic systems actually improves act/defer decision-making via rank correlation and equivalence testing.

Theoretical analysis connecting Föllmer processes to DDPM sampling via reverse SDE discretization and implications for sampling error.