Topic: Research
1414 stories · page 9 of 29
- Latent-Foresight trains world model tokenizer and predictor jointly
- JevSpawn paper proposes parallel action spawning for LLM agents
- Honeycomb video world model keeps scene memory at constant size
- Google Research unveils TEE-based federated learning with verifiable privacy
- Distillation study finds on-policy rollouts matter less than KL direction and learning rate
- CorrGRPO rebalances multi-reward GRPO with Pearson correlations
- Argo-Bench tests data agents on a 235-table, 7.5-billion-row warehouse
- World Observer video world model keeps out-of-view objects evolving
- VGBench tests whether audio LLMs stay silent when a bystander speaks
- SILSA replaces voxel tokens with slice latents for 3D generation
- Sharpening Tax paper: RL post-training costs agents solution coverage
- ServiceNow CoreAI details AutoSynthData, synthetic training tasks for enterprise agents
- SECRET: training-free method curbs audio-visual LLM hallucinations
- Rationales barely raise QA accuracy but sway verifiers, DeepSeek tests show
- PoS framework keeps explicit belief states for long-horizon LLM agents
- OneStreamer streaming video LLM leads eight benchmarks it was compared on
- Multimodal Flow models text and images in one continuous space, MF-1 released
- Microsoft Research ML pipeline forecasts space-weather risk to 66,935 US substations
- InterEvolve lets an LLM evolve reward programs for humanoid tasks
- Import AI 474: Levin's Platonic mind paper, Google TPUs in space, Zhipu's Infra Agent
- AlphaGo veteran Thore Graepel argues LLMs don't reason
- AgSpec uses retrieval-based drafting to speed up coding agents by up to 4.76x
- Adam drifts far from natural gradient descent under ill-conditioning, paper finds
- ActiveSaddler adapts training scenarios to improve agent harnesses
- 4Director controls video world models with rigid 3D object meshes
- VoxPolyMem gives agents long-term memory for multi-party spoken talk
- TabFM paper claims zero-shot tabular model beats tuned AutoML
- SaveRouter trains LLM routers on about 33 to 41% of the feedback
- Sage formalizes math into Lean 4 and cuts answer leakage to 2.7%
- Q&D trains agents to ask for information users never requested
- PMOPD protects task subspaces in multi-teacher distillation, beating MOPD on Qwen2.5-7B and Llama-3.1-8B
- Paper reproduces OpenAI-Hugging Face incident behaviors with public models
- Paper: LLMs are General Asynchronous Agents, shown with Qwen 3.x
- Livenerf benchmark tracks whether Claude Opus 5.5 gets worse after launch
- Hugging Face launches Open TTS Leaderboard for open-source speech models
- GPT-5.5 flags a planted negative result in just 2 of 200 reports
- GapFT fine-tunes on the Pass@K-Pass@1 gap to lift single-sample accuracy
- EngiWorld benchmark: best of seven frontier models scores only 44.3 on engineering software tasks
- Context Language Models let AI models manage their own context as a file
- Chunked KV-cache compression creates periodic weak spots, paper finds
- Alignment Forecasting paper predicts misalignment before fine-tuning, with a 5,000-question benchmark
- WideSWE benchmark: best coding agent solves 42.50% of multi-repo tasks
- WhiteMatter lets every Transformer layer reuse any depth's KV cache
- VisionHOPE proposes a self-modifying visual backbone with five coupled memories
- UMM-Reflection trains BAGEL to fix its own images with RL, +12.05 on GenEval
- TT-VidT video pretraining leads four benchmarks with fewer FLOPs
- Structured residual connections cut Diffusion Transformer training iterations by up to 1.73 times
- SCOPD self-distillation narrows the accuracy loss from visual token pruning
- RenderRank reranks documents as images, cutting input tokens 16.5-35.5%
- Relic turns multi-agent coding failures into persistent executable protocols