research
35 posts tagged research.
- On-policy distillation may need fewer prompts and better absorption
- On-policy distillation may be pruning tails, not teaching
- Reasoning Without Tokens: What Soft Latent Thinking Actually Changes
- Optimizers Are Becoming Systems Choices, Not AdamW Replacements
- SwarmWorld makes a case for agents that coordinate through artifacts
- Score models as reusable priors for detection
- A Newton Method That Hits O(1/k³) With One Linear Solve Per Step
- TSN4PI Wants to Predict Where Your Politics Are Headed
- LP-NAS puts a linear program inside differentiable architecture search
- What Six Years of TrustNLP Papers Say About Where AI Safety Research Actually Went
- Agnostic PAC learning gets its optimal bound
- Sparse Weight Decomposition makes circuit extraction less expensive
- Teaching Models Formal Logic Before Words: What Logic-PPT Actually Shows
- GradCuit optimizes the reasoning state, not the model
- LiveMem reframes long-context memory as state continuity
- On-policy imitation helps when the student is smaller than the expert
- ReToken Makes Visual Retrieval a One-Token Routing Problem
- The Price of Monoculture: What Happens to Writing When Everyone Uses the Same Model
- MODUS brings any-to-any multimodal modeling to decoder-only systems
- πR² makes robot policies react inside the action chunk
- Hyperball optimizers still need learning-rate discipline
- Quantum Spectral Models: Encoding a Matrix's Structure Into the Circuit Itself
- Barzilai-Borwein’s superlinear convergence problem
- Maskability Index makes prompt choice less vibes-based
- Sarcasm detection needs the mismatch, not just the meme
- RoboTTT treats robot memory as trainable state, not a longer prompt
- Transformer circuits may be lower-dimensional than they look
- Transformer Theory Moves From Can Represent to Can Learn
- PHINN-EEG’s dream detection claim is a proposal, not a result
- Co-LMLM Puts Facts in a Database, Not the Weights
- ELSA3D routes language to the right 3D scale
- Graph attention needs the graph’s spectrum, not an average filter
- C2R targets the hidden mess inside sparse autoencoder features
- DiScoFormer treats density and score estimation as one reusable job
- Tapered Language Models and the Free Lever in Layer Width