Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 32 · 2026
Aug 03 – Aug 09
16 papers, repos and datasets selected this week.
Papers
Picked Aug 7
- PaperTowards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes→
Finds early unification beats late/sequential training, and efficient recipes reach strong generation with only 5% of the compute budget.
- PaperWorldCycle: Self-Verifiable Reinforcement Learning for Long-Horizon Video World Models→
Reduces state-returning drift by up to 44% and improves composite-action accuracy by nearly 4× over the base model.
- PaperDAPD: Dual-Anchored Policy Distillation→
Beats OPSD by +2.00 avg on Qwen3-4B across 6 tasks; gains persist across scale: +2.69 at 4B and +2.78 at 32B.
Picked Aug 6
- PaperDeferred Exposure of Future Trajectories for Verifiable Reasoning in Autonomous Driving VLMs→
On AD-MCQ, DEFT-RLVR lifts accuracy from 28.1% to 77.9% on Qwen3-VL-8B and from 34.0% to 82.2% on Qwen3.5-4B.
- PaperDistillAlign: Coordinating Mode Covering and Mode Seeking in Autoregressive Video Distillation→
Best VBench 84.83 with 0.69 coverage; even the 1.3B-teacher variant scores 84.54, beating Wan-14B baselines like Causal-Forcing at 84.38.
Picked Aug 5
- PaperLongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks→
Boosts Qwen 3.7-Plus from 51.8% to 80.7% on WeaveBench, 69.7% to 77.2% on Terminal-Bench 2.1, and 2.8% to 8.3% on OSWorld 2.0.
- PaperDiffusionGemma Technical Report→
It converts Gemma 4 into a 256-token block discrete diffuser, reaching ~1,500 tokens/s on one H100 while preserving autoregressive and multimodal capabilities.
Picked Aug 4
- PaperScaling Properties of Text Conditioning in Visual Generation→
It links diffusion loss to caption informativeness, then uses image-grounded JSON prompts plus an LLM prompter.
- PaperSearch Beyond What Can Be Taught: Evolving the Knowledge Boundary in Agentic Visual Generation→
It co-trains a generator-specific gate that learns when retrieval helps, not blind search.
Repos
Picked Aug 6
- Repopatchy631/time-to-first-token→
Unlike benchmark-only serving repos, it delivers a 10-week build roadmap with reproducible commands and tests past 1000 concurrent OpenAI-compatible requests.
Picked Aug 5
- Repoinspatio/querysplat→
Unlike most 3DGS repos, it predicts scenes feed-forward from image folders and exports cameras and depth, with optional test-time optimization.
Picked Aug 4
- RepoFareedKhan-dev/kimi-k3-in-c→
Runs a 2.78T-parameter MoE on one CPU in 8.24GB RAM by streaming 1.56TB weights from disk and loading only routed experts.
Datasets
Picked Aug 7
- Datasetzai-org/AISE-Bench→
Human-verified academic KG benchmark includes gold multi-step API traces and grounded answers.
Picked Aug 6
- Datasetkylefengkfeng209/college-majors-2026-earnings-debt-jobs-ai→
It links 228k institution-by-major-by-credential programs to earnings, debt, job outlook, and AI exposure.
Picked Aug 5
- Datasetdatamatastudios/ai-requirements-index→
Makes AI hiring demand measurable over time across roles and seniority using a consistent listing-level share metric.
Picked Aug 4
- Datasetpsk/vaani-asr→
It pairs 172k quality-checked clips with native-script transcripts across 12 Indian languages and speaker-region metadata.