Only meaningful ML signals. Papers, repos & datasets. Selected, not collected. Home of The Weekly ML Signal → https://teapaw.org/t.co/wUFGgFytuv
Joined April 2026
- Tweets647
- Following22
- Followers185
- Likes146
Weekly Top · Nov 02 – Nov 08
Follow for next week’s top
📄 Paper: Proving and Improving the Platonic Representation Hypothesis with JEPAs
It proves cross-modal JEPA encoders align up to rotation and beats 72B LMs with 42x fewer samples.
💻 Repo: hyz317/HairDeck
Unlike most single-image 3D tools, it outputs hair cards, spring-bone rigs, and GLB/VRM exports together.
📊 Dataset: Axion1-0/axion-training-data
Unlike mostly evaluation benchmarks, it releases 18,363 full agent-training trajectories plus 484 RL tasks with verified tool-use.
Full issue: papersdatacode.com/issues/12…
ALT ML Weekly Top, Nov 02 – Nov 08: Real-world agent evaluation. Long-horizon benchmarks, tool-use trajectories, and robot action labels.
📋 ML Weekly Recap · Nov 02 – Nov 08
⚡ Trends
▸ Benchmarks increasingly target long-horizon, real-world agent and robotics evaluation
▸ Unified multimodal models replace modality-specific stacks with shared representations or schemas
▸ Inference efficiency advances emphasize streaming, speculative decoding, and adaptive runtime routing
🧭 TL;DR
📄 Proving and Improving the Platonic Representation Hypothesis with JEPAs
Randall Balestriero, Yann LeCun
Proves cross-modal alignment and beats giant LMs with 42x fewer samples
📄 arxiv.org/abs/2610.39302
⭐ Mika
Practical on-device Android automation across apps without APIs or cloud
⭐ github.com/wangzan101/Mika
💡 Agents are moving from toy tasks toward real-world, verifiable workflows.
→ t.me/papersdatacode_digests
ALT ML Weekly Recap · Nov 02 – Nov 08. Long-horizon agent benchmarks; Unified multimodal models; Inference efficiency.
💻 GenLang-V/GenDB-V
Builds verified Rust query engines for a schema or hot query template by translating SQL to a fixed spec, then generating code and proofs checked by Verus.
Key points:
• Up to 61.3× over DuckDB and 28.3× over Umbra on SEC-EDGAR
• 51.6× over DuckDB and 22.0× over Umbra on TPC-H 22 queries
• Verus proves correctness, memory safety, arithmetic safety, no data races
Shows AI can synthesize database engines that are both formally verified and substantially faster than mature systems on targeted workloads.
⭐ 15 stars
ALT Pipeline diagram showing GenDB-V synthesis, verification, and trust boundaries
📄 ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents
by Guanting Dong (@kakakbibibi), Zhichao Sun et al.
A trace-first framework that replaces data-centric environment synthesis with execution-witnessed tool graphs and replay-verified task generation across environments.
Key points:
• Qwen3.6-35B-A3B improves from 23.0 to 41.0 avg; 43.2 with RL
• GLM-5.2 improves from 44.6 to 56.6 avg across 7 benchmarks
• On Pro benchmarks, 7 frontier models drop by up to 38.2 points
Real execution-witnessed tool composition is a stronger primitive for training and benchmarking long-horizon agents than synthetic, LLM-judged environment graphs.
ALT Two-stage pipeline diagram showing the full method for task synthesis
💻 inclusionAI/IDM
Runs 1-step or 2-step text-to-image sampling from one checkpoint and improves text-heavy prompts using EMA feature-moment matching plus internal DMD.
Key points:
• 94.25% LongText-CN and 90.20% LongText-EN at 2 NFE
• 0.25s at 1 NFE and 0.30s at 2 NFE on one A100, BF16
• EMA moment loss adds +0.41 LongText-EN over internal DMD alone
Shows that a generator's own EMA features can replace an external encoder for moment matching in few-step text-to-image distillation.
⭐ 240 stars
ALT Method pipeline showing DMD loss and feature-moment loss architecture
📄 Route Anything: A Study of Model and Policy Routing for General-Purpose AI Systems
by Pengfei Zhou, Yixing Ma et al.
Per-input routing across heterogeneous models and robot policies, optimized for quality-cost tradeoffs and trained on candidate outcome prediction.
Key points:
• Oracle routing beats best fixed by 5.4 pts across 24 benchmarks
• Learned routers gain 2.1 pts on video generation vs best fixed
• Robot routing reaches 71.6% vs 70.5% best single policy with 24% less time
Shows that much of the gain in general-purpose AI systems comes from input-level orchestration of specialized candidates, not from picking one globally best model.
ALT Shows router pipeline plus headline routing gains and efficiency, core contribution
💻 wangzan101/Mika
by @wangzan101
Operates Android apps and websites from chat prompts, using a background screen so you can keep using the phone while it works.
Key points:
• No Mika server; screen data goes only to the configured model provider
• Works with any app via screen understanding, not per-app APIs
• Compatibility self-check runs 12 checks in about a minute
Shows that a practical cross-app Android agent can run locally on the phone while staying usable during execution.
⭐ 6.6k stars
ALT Screenshot showing AI agent operating phone app, demonstrating product capability
📄 Library of Layers (LoL): Adaptively Ordering Transformer Layers at Runtime
by Abdellah Rahmani (@arahmani_AR), Sean McLeish (@SeanMcleish) et al.
A router selects the next block from a shared layer library at each depth, so execution order changes per input instead of staying fixed.
Key points:
• Safety probe on sequence-level routing features reaches 0.95 AUROC
• Maze strategy control hits 99.79-100% correct forced-strategy traces
• Sudoku forced routing yields 99.5-100% requested-strategy accuracy
Shows that layer order can be a learned, controllable part of transformer computation rather than a fixed architectural choice.
ALT Detailed method diagram showing layer architecture, strided attention, and router mechanism
📄 CUA-Master: Scaling Computer-use Agents through Open-World Recursive Improvements
by Xinlei Yu, Gen Li et al.
Open-world video workflows become seed tasks, then reusable primitives synthesize controllable GUI/CLI tasks with rollback-based data collection.
Key points:
• 80.2 avg on OSWorld Verified, competitive with top closed-source agents
• 43.9 avg on WeaveBench, above UI-Mate-27B's 18.3
• 25.0 full-pass on Agents' Last Exam, above Qwen-CUA's 5.3
Shows open-world seeded, recursively targeted data generation scales computer-use agents better than closed self-improvement alone.
ALT Detailed pipeline diagram showing the recursive flywheel method and components
📄 Wave Forcing: Towards Speed-of-Light Streaming Video Generation
by Yijie Jin, Yuxuan Liu et al.
Pipelines chunk denoising and history refresh across replicas, using causal access to cleaner in-flight chunks to sustain continuous video output.
Key points:
• 134.8 end-to-end FPS and 147.6 DiT FPS on 8 H200s with 4-step WF
• 5-step WF: 81.64 VBench Total vs 81.15 for Rolling Forcing
• Long-video Dynamic 91.45 vs 36.63 for Rolling Forcing; Subject 97.10 vs 92.75
Shows streaming video generation can reach near-real-time throughput without giving up long-horizon motion or subject consistency.
ALT Method diagram showing wavefront denoising, causal K/V delivery, and continuous output pipeline
📄 On the SPOT: Enhancing Speculative Decoding for Diffusion Language Models via Commit-Order Prediction
by Eliezer Shlomi, Eliron Rahimi et al.
A learned commit-order drafter for DLMs that reuses stale token predictions and verifies drafted state trees exactly with shared-noise sampling.
Key points:
• Up to 3.4x throughput vs standard decoding on SDAR-8B
• Uses 16-36% fewer target forwards than Spiffy on SDAR-8B
• SPOT-Sets adds 59% avg throughput over threshold decoding alone
For diffusion LM decoding, predicting where to commit next matters more than predicting the token itself.
ALT Method diagram comparing target decoding vs SPOT tree-based parallel decoding approach
📊 physical-engineerings-last-exam by google
Contains realistic engineering task workspaces across 4 disciplines, with project files and tools for agents to produce engineering deliverables.
Key points:
• 4 engineering disciplines
• Includes drawings, specs, spreadsheets, and solver scripts
• Uses CAD/BIM or simulation tools in the workspace
Useful for benchmarking agents on long-horizon, multi-tool engineering task execution.
📥 1.3k downloads
ALT Pipeline diagram showing the adversarial creator-solver co-curation loop method
📄 Latent Reasoning in Continuous Space for Unified Multimodal Models
by Byungwoo Jeon (@rootyjeon), Yoonwoo Jeong et al.
Replaces discrete token reasoning with continuous latent reasoning for UMMs, then self-improves it via information-gain RL tied to image prediction.
Key points:
• GenEval overall 0.88 vs BAGEL 0.81 and Uni-CoT 0.83
• DPG-Bench overall 90.04 vs BAGEL 85.07
• OneIG-Bench overall 0.866 vs BAGEL 0.764 and Uni-CoT 0.790
Shows that aligning reasoning with the continuous image-generation space unlocks stronger controllability and real test-time scaling in unified multimodal models.
ALT Two-stage pipeline diagram showing SFT and Self-Evolving RL method architecture
📊 RoboAtom by InternRobotics
Contains time-aligned robot trajectory segments from 12 source datasets, each with frame bounds, skill label, and action text for segmentation tasks.
Key points:
• 12 source datasets in the concatenated annotation set
• 29 predefined skill labels plus unlabeled spans
• 500 benchmark trajectories with 4,018 GT segments and 721 videos
Useful for training and evaluating models that segment robot videos into atomic skills and generate action descriptions.
📥 431 downloads
ALT Composite overview with dataset stats, benchmark coverage, annotations, and comparative results
📄 VibeVideoBench: Evaluating Multimodal Agents on Code-Driven Video Generation
by Muxi Diao, Wenbin Xu et al.
Evaluates general agents on making complete videos via code in a sandbox, with hierarchical checks for spec compliance, task requirements, and quality.
Key points:
• Best strict success: 76.3% of tasks by Claude Opus 5.5
• GPT-6 Astra ranked #1 on delivered-video L3 overall with 4.126
• Judge system ranking matched human preferences with ρ=0.98
For code-driven video agents, planning and delivery reliability are as important as raw content quality.
ALT Full pipeline diagram showing sources, curation, rubrics, rollout, grading, and judging.
📄 Proving and Improving the Platonic Representation Hypothesis with JEPAs
by Randall Balestriero (@randall_balestr), Yann LeCun
Good-view JEPAs use paired observations of one scene, not token continuation, so separately trained encoders align up to one rotation.
Key points:
• 109M text JEPA gets .184 vs Qwen2.5-72B .154 on DINOv2 alignment
• Matches next-token prediction with 21.6M vs 900M samples, 42x fewer
• Pair-free alignment hits 64% top-5 on ImageNet-100 vs prior LM best 56%
Cross-modal platonic representations depend mainly on the positive-view structure, not on scaling next-token prediction alone.
ALT Combines method diagram with scaling curve showing core alignment claim vs baselines
Weekly Top · Oct 26 – Nov 01
Follow for next week’s top
📄 Paper: Hy-Nomas: Towards Free-form Interactive World Generation
Unlike other world models, it supports multi-turn free-form video interaction by separating persistent scene memory from local action commands.
💻 Repo: CuboiLeo/SuperMimic-Reproduced
Unlike most control repos, it fully reproduces SuperMimic on Unitree G1, matching 0.82 LAFAN1 in 1.66 hours on one RTX 5090.
📊 Dataset: Name That LLM Open Pool: 255k Passages, 18 LLMs
Unlike most attribution datasets, it provides 254,561 de-duplicated open passages labeled across 18 LLMs with provenance and leakage flags.
ALT ML Weekly Top, Oct 26 – Nov 01: From worlds to grounding. Free-form video interaction, fast box recovery, and multilingual tasks.
📋 ML Weekly Recap · Oct 26 – Nov 01
⚡ Trends
▸ World models are becoming interactive, rewindable, and language-conditioned for control
▸ Agent research is shifting toward capability-based training and harness reliability evaluation
▸ Benchmarks increasingly target grounded, localized, and long-context multimodal understanding failures
🧭 TL;DR
📄 The Trinity:Capability-Driven Recursive Self-Improvement for Multimodal Agents
Weidong Tang, Ruiheng Wang, Jingxuan Zhang et al.
Capability-organized recursive self-improvement delivers large gains and broad multimodal generalization.
📄 arxiv.org/abs/2610.36077
⭐ WeVisOne-VL
Unified perception repo covers many vision tasks with strong open performance.
⭐ github.com/Tencent/WeVisOne-…
💡 ML is shifting toward unified, reusable agents trained across diverse environments.
→ t.me/papersdatacode_digests
ALT ML Weekly Recap · Oct 26 – Nov 01. Interactive world models; Capability-based agents; Grounded multimodal benchmarks.
📄 Can Discrete Tokens Serve as Semantically Lossless Representations of Visual Signals?
by Jialv Zou, Xiaotong Li et al.
A bit-budget-first discrete visual tokenizer that uses parallel codebooks on complementary full-feature views, rather than residual or channel-split quantization.
Key points:
• 65.3 avg on 8 benchmarks vs 64.6 for continuous QwenViT
• 22.8% higher 9-metric average than UniAR
• 59.2% higher 9-metric average than Emu3-Chat
Discrete visual tokens can match continuous visual representations when added bits remain semantically informative.
ALT Training pipeline and FVQ method diagram showing the core architecture