The Archive

Decade-long study of Android malware detector adversarial robustness under temporal concept drift across deployment scenarios.

Google Embeddings 2 outperforms five open-source dense retrieval models on BEIR and RAG benchmarks but faces latency tradeoff.

Entity-centric latent patch memory for multi-shot video generation maintains character consistency without full-frame overhead.

DiLaDiff uses latent diffusion and consistency model distillation to improve token correlation in masked diffusion language models.

Preisach Attention Layer applies hysteresis operator as O(1)-depth attention replacement, achieving Turing-completeness in single-layer Transformers.

Fine-tuning LLMs for entity resolution in KYC via structure-guided matching across naming conventions and scripts.

MetaEvaluator: meta-learning framework for label-free, cost-effective evaluation of unseen models across architectures.

Scheduling algorithms for aircraft disassembly optimization with task precedence and certification constraints.

Scaling laws for sparse-activation neural networks show double-descent and asymmetric loss dynamics.

Co-ReAct integrates step-level rubrics to guide ReAct agent reasoning in multi-step search tasks.

Study of training sample requirements for neural network inverse kinematics in robotic manipulators.

PushBench evaluates quantitative goal persistence in long-horizon LLM agents via work-unit completion.

that meme on the chatgpt subreddit is so spot on ngl. even when you have requirements locked down managing the stack gets so weird. claude is an absolute beast at backend logic, teh reasoning depth is just insane now.the real mess starts when u try to scale past a basic landing page. forcing a single chat window to track complex UI layouts on top of everything just cooks the token limit and causes massive code drift. i ended up completely separating my enviornment to stop fighting the bottleneck. now i just let claude handle pure data pipelines, dump states into a quick db, and let stitch tak...

HARNESS-LM distills large embedding models into compact SLMs for low-latency sponsored search.

Hybrid DP-CP approach for partial shop scheduling combines dynamic programming with constraint propagation.

User reports Claude Max outperforms ChatGPT Pro on accounting/taxation/legal tasks despite higher token costs; anecdotal quality comparison in Indian context.

SupraLabs released Supra-50M, a 50M-parameter Llama-style language model trained on 20B educational tokens with competitive benchmark performance.

Analysis of 100+ sequential RL training pipelines shows salient feature-driven generalization and goal persistence.

MARS improves model evaluation by weighting ranks with performance margins instead of discarding magnitude differences in Critical Difference diagrams.

ARMS auto-generates dense reward shaping signals for multi-agent RL via trajectory ranking without task-specific retraining.

PathNavigate applies training-free agentic VQA to whole-slide pathology images using surprise-guided navigation and memory caching.

Theoretical study of why low-dimensional embeddings scale to trillions of retrieval targets via maximal-margin analysis.

Goal-conditioned RL agent leverages parallel all-goals learning by jointly predicting values and actions for every goal simultaneously.

RA-DCA proposes randomized active-set optimization for nonsmooth max-structured DC programs with directional stationarity guarantees.

Method addresses visual artifacts in dimensionality reduction by detecting and splitting ambiguous instances across multiple neighborhoods.

Most everyone focuses on Claude, the Constitutional AI Safety Research. However, I believe that the most practical impact from anything Anthropic has released to date may have been MCP. Given that MCP is a model-agnostic platform that is open-source, it allows developers who are not utilizing Claude to utilize it as well. Both OpenAI and Google are utilizing MCP. As such, MCP is being developed into the de-facto industry standard for connecting tools within artificial intelligence. I also find MCP shifts the bottleneck. Historically, getting an LLM to become smarter was the difficul...

Precise applies RL post-training to flow-matching models by designing SDE-consistent stochastic samplers that respect reverse ODE dynamics.

NHODE combines Hamiltonian neural networks with neural ODEs to learn partially observed dynamical systems without full state access.

DrawVideo generates long-form video from storyboard sketches, decomposing sequences into independently controllable shots via sketch/appearance/motion prompts.