model-training
91 posts tagged model-training.
- AutoScaffold trains models with hints, then removes the hints
- FlowAlign treats alignment as target-space geometry, not just matching pairs
- What 'AI accelerating AI research' actually looks like inside a lab
- A2ODA is a small optimizer idea with a practical training lesson
- An Independent Replication of Direct Nash Optimization, and What Its Ablations Actually Prove
- Monofacts may be the data ingredient behind AI scientific discovery
- Sonar-MoE cuts MoE fine-tuning memory by moving optimizer steps into backward
- LARC moves image reasoning out of token space
- When an AI vendor refuses model-training work
- SpecSlot splits image tokens by what needs global attention
- MuonGS relaxes weight-norm control without giving up training speed
- Reading Evolution Strategies as a Center of Mass, Not a Gradient
- AdaBPE makes tokenization care about LaTeX
- DexPort makes robot hand data collection less robotic
- Nanbeige4.3-33B is really a small-model bet
- Speech-only conversation models move the hard part to turn-taking
- Adaptive Depth in Recommenders: What LoopRec Changes
- LongZero trains long-context reasoning without outside documents
- Sparse RLHF makes private preference data less impossible
- TS2Vec-FM shows small convolutional time-series models still have legs
- WildAWM trains computer-use agents in synthetic worlds
- Pixel diffusion gets a smarter prediction target
- TorchNeuron points at the real Trainium bottleneck: developer friction
- ORL Makes Diffusion LLMs Look Practical for Low-Latency Inference
- Indic pretraining is becoming a data pipeline problem
- CoPE trains agents to forecast their own future actions
- Diffusion Models Are Starting to Generate Robot Policies, Not Just Robot Videos
- Back-translation beats more data in low-resource translation
- Robot demos get more useful when the model knows how messy they are
- Tabular regression still punishes lazy early stopping
- Nvidia Buying an 'Open' Model Lab Is About Compute, Not Openness
- SGUID says most skills are not worth distilling
- Tokenizer compression is not the same as multilingual quality
- GPU scheduling is where AI infrastructure gets real
- The missing model is becoming a product requirement
- JOLT makes self-distillation follow the student’s path
- Multi-teacher distillation should copy the delta, not the whole teacher
- IdeaAnchor Trains Models on the Shape of Research Ideas
- Sherpa trains tutors on learning outcomes, not lesson vibes
- Diffusion models may be sketching meaning before pixels
- Paradee makes single-voice TTS small enough for a laptop
- FERPO avoids action-gradient trust in continuous-control RL
- SILSA compresses 3D generation by slicing the shape, not voxelizing everything
- What HC-DLM Fixes About Diffusion Language Models
- Looped Mixture of Experts: Two Efficiency Tricks That Finally Stack
- Wild AI Web Text Is Starting to Poison Pretraining
- NoPE attention is not position-blind if local layers carry recency
- Robots Need Correction Loops, Not Just Better Planners
- Foil's Recipe for Looped MoE: Flatten the Experts, Untie the Attention
- For Linear ViTs, copy the MLP and distill the attention
- KV-streams make long-agent training cheaper by not throwing memory away
- Telescopic Language Models: one training run that serves every compute budget
- X-Reset turns human hand demos into better robot starting points
- Free local AI models are not charity
- Clifford-VAE puts pixels into symbolic memory space
- UECR-GRPO treats the teacher as evidence, not an oracle
- Critical-State RL trains the one agent call that actually matters
- onPanda turns alignment feedback into token-level steering
- Where Harness Self-Improvement Actually Stands in Late 2026
- Membership Inference Gets an Entropy Correction: Reading the ETD Paper
- Mini-AGI makes local training the interesting part
- Softmax attention needs an off switch
- Microsoft’s AI scraping memo points to a data supply chain problem
- Paint-Anything makes hex colors a first-class diffusion control
- Double descent as an implicit regularization story
- What Actually Makes a Tokeniser Good: Search Beats Objective
- Distillation needs calibration when the teacher is biased
- Bellman Policy Optimization cuts one moving part from RLVR
- Federated learning gets more practical when privacy and timing are treated together
- OptiFlow treats offline RL policy learning as sample matching
- Garry Tan’s distillation argument is really about AI capability access
- The useful part of calling Nvidia AI’s central bank
- CRISPR screens need learned experiment pickers, not bigger chatbots
- Recursive Self-Improvement Has a Roadmap Now, and Most of It Isn't Built
- Layer-selective unlearning aims at the parts of a model that remember
- On-policy distillation may be pruning tails, not teaching
- Optimizers Are Becoming Systems Choices, Not AdamW Replacements
- Reward choice changes how LLM forecasters are wrong
- Alignment baked into pretraining, not bolted on later
- Selective Trust: Why RAG Systems Should Learn When to Ignore Their Own Context
- MindForge trains coding agents on blank-repo software work
- Hyperball optimizers still need learning-rate discipline
- MIRROR trains vision models by making each modality teach the others
- OPD2 tries to distill reasoning by subtracting the base model
- RL post-training as procedure compression, not just skill amplification
- Timestamp drift is the quiet ASR failure that breaks real workflows
- DemoPSD Treats Teacher Disagreement as a Training Signal
- RLVR needs a taste model, not just a grader
- Synthetic QA Has a Selection Problem Before It Has a Training Problem
- Autodata Turns Synthetic Data Generation Into an Agent You Train
- Self-distillation can make models better on the first try and worse on the fifth