ai-research
195 posts tagged ai-research.
- AutoScaffold trains models with hints, then removes the hints
- B3 speeds up diffusion language models by overlapping the slow part
- Contrastive Discovery: Why Science Agents Should Weigh Rival Explanations, Not Just Reject One
- HiVe-Mem and the Case for Two-Dimensional Agent Memory
- MaxIM treats agent memory as a trained compression problem
- A Generalist Model Beat Specialist Robot Policies Without a Single Demo
- FiLSim pushes graph similarity past toy graphs
- FlowAlign treats alignment as target-space geometry, not just matching pairs
- FP4 KV Cache Done Right: What BitDecoding-FP4 Changes for Long-Context Serving
- Grasp-TAGS picks the grasp before the robot commits
- Protein model steering still breaks at the decoder
- What Microsoft Shutting Down AI Frontiers Says About Industrial Research
- What 'AI accelerating AI research' actually looks like inside a lab
- Does Looking Inside a Model Actually Make It Safer?
- MAEMs turn residual stream directions into testable text
- MedGraph treats medical memory as patient state, not chat history
- Scientist World Turns Research Papers Into Runnable Environments
- Animal ID Models Need Evidence, Not Just Retrieval Scores
- HALO Trains Humanoids in a Thousand Simulated Homes, Then Ships to the Real One
- Richer Context Does Not Automatically Fix LLM Social Agents
- A2ODA is a small optimizer idea with a practical training lesson
- Cognitive symbionts shift the agent question back to the human
- Fixed Personas Break on Context in Synthetic User Tests
- LoopVQ and the Hidden Memory Cost of Looped Transformers
- Monofacts may be the data ingredient behind AI scientific discovery
- Why VLM robot reward models hallucinate progress, and a prompt fix that works
- KLM Wants AI to Show Its Evidence. Its Own Negative Results Show Why That's Hard.
- When the Chain of Thought Doesn't Actually Drive the Answer
- Zyphra's tree-structured attention aims at million-token-plus context
- Factorio Peaceful Mode Is a Useful Agent Test, Not a Finish Line
- AI clutter suppression makes 5G sensing less brittle
- Neural Thickets May Be an Inference Trick, Not a Weight-Space Discovery
- SpecSlot splits image tokens by what needs global attention
- TextSeal2 keeps watermarks alive under speculative decoding
- The Economics of Proactive AI: When a Single 1B Decoder Beats a Two-Model Pipeline
- Project ProtoStar puts the boring data layer under science AI
- GAS-PFN brings road topology into zero-shot traffic forecasting
- Hexels push 3D reconstruction toward fewer primitives
- Latent world models are passing tasks without learning the world
- MuonGS relaxes weight-norm control without giving up training speed
- What LongChang's Altered-Physics Test Reveals About LLM Curiosity
- Wilson loops give relational attention a harder rule
- Clinical notes as an early warning layer for hospital mental health risk
- dSpark and the Order Problem in Diffusion Language Models
- FLORE shows why humanoids need force, not just motion
- OpenAI’s science claims need verification, not applause
- 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
- SemantEq makes formal-math evals less easy to game
- When an LLM's 'True' and 'False' Don't Sum to One
- CausalCoT treats chain-of-thought as a dependency graph
- Pruning reasoning candidates by optimizing pass@k directly
- Adaptive Depth in Recommenders: What LoopRec Changes
- BatchBench puts biomedical vision models through the real failure mode
- Quant-Arena Tests Whether LLMs Can Actually Build Trading Strategies
- Sparse RLHF makes private preference data less impossible
- Why the Best Solver Model Is Rarely the Best Environment Author
- Adaptive MTP makes speculative decoding less wasteful
- NVIDIA’s JCA turns generated video into rig motion
- The Teacher That Scores Worse Than It Plays: Fixing Agentic Distillation
- What Arabic Dialect Math Problems Reveal About Model Reasoning
- AfriBED makes speech bias measurable in Bemba and Nyanja
- WeAct Turns LoRA Adapters Into Something an Agent Can Pick Mid-Task
- An AI Agent That Reproduces Cancer Studies and Checks Where They Wobble
- Pixel diffusion gets a smarter prediction target
- LeanGraphSearch makes theorem provers follow the proof graph
- PathFM Treats Forecasts as Paths, Not Quantile Spreads
- Indic pretraining is becoming a data pipeline problem
- MLExpert and the Case for Training Agents on Whole Experiments, Not Single Turns
- OpenAI’s GPT-Rosalind plan puts life-science AI inside academic labs
- Google DeepMind’s audit separates code volume from research speed
- HiRe-MoMa fine-tunes robots by correcting policies instead of replacing them
- PerSo splits persuasion into reading people and choosing the next move
- The Case for Making Research Agents Forget Their History
- When Renaming a Tool Breaks Your Agent: What CIPR Fixes
- DoubtLess uses a small model to make a big one stop overthinking
- Temperature Should Read the Room
- AI Cancer Claims Need Milestones, Not Cigarette Theater
- Can vision-language models actually spot persuasion in memes?
- Reading 4,000 Comic Books at Once: What the PAGE Framework Actually Does
- Tabular regression still punishes lazy early stopping
- The Same Attention Head Does Different Jobs Depending on the Task
- Tool retrieval should know when the tool is broken
- SGUID says most skills are not worth distilling
- Tokenizer compression is not the same as multilingual quality
- ViSkill and the case for keeping agent memory visual
- What SpaceCast-Bench reveals about the gap between seeing space and predicting it
- When One Input Has Many Right Answers: Bi-FORK and the Bifurcation Problem in ML Surrogates
- Multi-teacher distillation should copy the delta, not the whole teacher
- OpenAI’s withdrawn math results are a reminder that AI math needs audit trails
- Robot policies break on synonyms, not just scenes
- Scoring LLMs When There Is No Right Answer: A Lesson From Survey Economics
- The Explorer and the Student: How ExpDis Splits RL Into Two Jobs
- When a hallucination enters the context, can the model talk itself out of it?
- IdeaAnchor Trains Models on the Shape of Research Ideas
- OpenAI's Math Proofs on GitHub: What the Lean Files Actually Show
- Recommendation systems need a throttle, not just a better ranker
- CLIFT turns web-agent judging into reusable verification
- Diffusion models may be sketching meaning before pixels
- Robot agents need indexed video memory, not bigger prompts
- The First Word a Model Writes Decides How Well It Reasons
- Running Anomaly Detection on the Satellite Instead of the Ground
- FERPO avoids action-gradient trust in continuous-control RL
- GALA makes Gaussian avatars lighter by distilling animation into shared blendshapes
- ScholarCatalyst: When Agentic Search Can't Beat a Plain Embedding
- SILSA compresses 3D generation by slicing the shape, not voxelizing everything
- Source learning makes RAG less forgetful
- What HC-DLM Fixes About Diffusion Language Models
- Looped Mixture of Experts: Two Efficiency Tricks That Finally Stack
- MatLoom makes generated materials editable, not just pretty
- The multilingual loophole in LLM unlearning
- When the Prompt's Phrasing Changes the Answer: SCAPO and Token-Level Credit in RLVR
- Wild AI Web Text Is Starting to Poison Pretraining
- Grounded Entity Biographies: A Different Fix for Long-Video Memory
- NoPE attention is not position-blind if local layers carry recency
- Not All Quantization Errors Are Equal: STEPQuant and the Memory Cost of Linear Attention
- Foil's Recipe for Looped MoE: Flatten the Experts, Untie the Attention
- For Linear ViTs, copy the MLP and distill the attention
- One uncertainty layer for graph neural networks
- Flow matching gives neural dissimilarity metrics one common frame
- Teaching a Reasoning Model to Know When It's Sure Cuts Its Token Bill
- What LLMs Miss About Haitian Creole, and Why Low-Resource Culture Breaks Evals
- Conspiracy Detection Needs Context, Not Just Better Keywords
- ExplorationBench and the Gap Between Recall and Real Discovery
- MISVO steers frozen language models at inference time
- Authorship verification works better as similarity, not classification
- Clifford-VAE puts pixels into symbolic memory space
- When a Model Gets Math Right but Reads the Same Problem Differently
- A Small Model That Steers a Bigger One's Reasoning
- Long-context models can miss the answer sitting behind nearby noise
- OOD generalization depends on exact mechanisms, not better fits
- Membership Inference Gets an Entropy Correction: Reading the ETD Paper
- OpenAI’s math advisory group is about claim control
- Softmax attention needs an off switch
- Decontamination reports are a promise, not a proof: the case for evaluation-side benchmarks
- A cipher win is not an eval without the working
- ComPO and the Case Against Optimizing the Loss You Wrote Down
- Double descent as an implicit regularization story
- Probabilistic Linear Explanations make interpretability more honest
- What Actually Makes a Tokeniser Good: Search Beats Objective
- When Rewording the Answer Key Reshuffles the Leaderboard
- Fuse tests the weak spot in AI social advice
- LACE compresses speech tokens one codec layer at a time
- ScienceBuddy and the case for training the harness before the model
- Bellman Policy Optimization cuts one moving part from RLVR
- LLM personas fail when opinions have to change
- OptiFlow treats offline RL policy learning as sample matching
- Slip Detection Is Where Robot Hands Stop Dropping Things
- The Skill Router You Already Have: Gavel Reads Routing From a Frozen LLM
- AI math systems are still optimizing for the wrong proof
- Conversational XAI works best when it is a control surface, not a chatbot
- CRISPR screens need learned experiment pickers, not bigger chatbots
- DataShifts turns distribution shift into an error-budget problem
- MindTopo Tests Whether Vision Models Grasp Topology, Not Just Distances
- Recursive Self-Improvement Has a Roadmap Now, and Most of It Isn't Built
- ConvMem Turns Long-Context Reasoning Into a Tree Instead of a Chain
- Geometry reasoning gets better when the model is not doing every job
- OpenAI training-data accusations need a provenance test
- SG-JEPA points world models at the encoder, not the simulator
- NOAH models patient records as timelines, not snapshots
- Procedural Graphs give agents a memory of what to do, not just what happened
- AlphaGenome Atlas turns genome variants into an AI lookup layer
- OpenAI Says Its Own Researchers Now Lean on Coding Agents. What Does the Data Actually Show?
- Claude, Fermat, and the real value of machine-checkable proofs
- Auxiliary views explain why diverse pre-training data works
- Verbal reinforcement learning is a feedback routing problem
- SWE-Prime argues coding agents need cleaner wins, not more wins
- WikiSkill and the Case for Giving Agents a Memory That Compounds
- Anatomy-informed networks put clinical constraints before more data
- Agent memory works better when skills are small and written in text
- AlphaEvolve nudges the matrix multiplication exponent lower
- AutoSR turns symbolic regression into a research-state search
- PIHF turns prompts into a versioned policy layer
- YOPO makes abstention cheaper by reading the model before it lies
- Alignment baked into pretraining, not bolted on later
- AutoDesign turns paper-to-poster into a harness the agent rewrites itself
- LLMs Know When to Back Off, But Still Guess Too Specifically
- OmniScientist argues that AI scientists need eyes, not just workflows
- CLAUDE.md bloat is a memory problem, not a prompt problem
- Concise answers can weaken reasoning in fused LLM training
- Two prepared policies may be the sweet spot for uncertain MDPs
- AI research agents can code, but they still can’t judge the work
- The Automation Ceiling Nobody Prices In: When Human Participation Is the Product
- No Best Harness: Automated Discovery Systems Fail to Generalize
- Uncertainty metrics should follow the loss, not the other way around
- Vision models are getting the scene right, but not always the gaze
- Leanstral 1.5 points at proofs as a workflow, not a stunt
- DiaLLM separates dialect understanding from dialect writing
- Language critiques are a better training signal than a score, if you can afford them
- LLMs brainstorm like synthesis machines, not researchers
- TraceLab shows coding agents are an infrastructure workload now
- Google's Paper Assistant Wants to Catch Your Math Errors Before a Reviewer Does
- Nash solvers have preferences when the value is identical
- AI math is less about genius than verification