Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 42 · 2026
Oct 12 – Oct 18
20 papers, repos and datasets selected this week.
Papers
Picked Oct 16
- PaperCortex: Consolidating Context into Parametric Memory→
Across 4B-120B models, Cortex is on the CR-RR Pareto frontier in most settings. After 50 sequential factual edits on 14B, Cortex-Structural gets 88.9% CR and 99.8% RR.
- PaperPlayer-1: Learning to Play from Synthetic Experience→
Uses synthetic RL across 16 verifiable games to learn transferable computer-use control, boosting MineDojo-300 success from 10.54% to 38.33%.
- PaperVideoJev: System One Model for Video Understanding and Interaction→
VideoJev-35B-A3B reaches 67.7 mean grounding mIoU, +10.8 over its backbone, with 3.1x speedup; QA stays competitive with Thinking mode while running over 100x faster.
Picked Oct 15
- PaperWorkRover: Scaling Long-Horizon Cowork Agent Training Via Recursive Environment Evolution→
Post-training Qwen3.5-35B-A3B gains +26.27 on JobBench, +29.32 on OfficeQA Pro, and +413.25 Elo on GDPval v2.
- PaperJeva: System One Models from First Principles→
Jeva-4B gets 92.05% on BANKING77 K=8, 0.00% permutation flips, 98.66 AUROC for nulls, and >20× cache speedup.
Picked Oct 14
- PaperMemento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks→
On ARC-AGI-3 it clears all 25 public games, reaches 100.0 mean RHAE, and uses 44% of human actions.
- PaperPhysEvo: Astra Can Act, Let It→
Across 42 RoboDojo tasks, it gets 62.0% success vs 47.17% for RoboDawn; on 8 hard Astra tasks, 55.0% vs 1.25% direct Astra.
Picked Oct 13
- PaperS-Space: Exploring Spatial Workspace in Multimodal Language Models→
It identifies a fixed linear 3D spatial subspace in MLLM hidden states and causally edits it.
- PaperDiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models→
Prompt-aggregated reverse-rollout guidance plus trust-region restoration matches baselines using 12x fewer GPU hours.
Picked Oct 12
- PaperAgentGarten: Code Worlds for Evolving Agents→
Agents learned from 4 rounds of interaction versus millions for a conventional RL counterpart; hide-and-seek behaviors like shelter building and ramp use emerged within a handful of games.
- PaperSelf-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight→
Across 10 agentic/reasoning tasks, SRD adds up to 24.2 points. In 2B with 98% all-failure groups, RLVR gets 0.0% success; +SRD reaches 60.6%.
Repos
Picked Oct 16
- RepoOptima-CityU/OpenArchEvo→
Searches executable PyTorch architecture programs beyond fixed NAS spaces, yielding spiking models with 31.7× lower estimated arithmetic energy.
Picked Oct 15
- Repoziplab/ATAS→
Adds explicit scene and event memory with validated world updates, targeting spatial revisit and event recapture beyond standard multi-shot generation.
Picked Oct 14
- Repophysicalvision/artisan→
Adds end-to-end generation plus a benchmark covering appearance, geometry, materials, collisions, stability, and mass.
Picked Oct 13
- Reporockmatrix-ai/YANchor→
Unlike most long-context repos, it pairs a recurrent 4B model with O(1) memory and H100 serving that sustains 512 sequences at 12,431 tok/s.
Picked Oct 12
- RepoNandhaKishorM/vegaml→
Unlike recent typed-decision repos, it adds local multimodal decisions with conformal sets and abstention.
Datasets
Picked Oct 15
- DatasetHelenPeng/ExeWorldBench→
Targets executable Blender world construction, scoring functional correctness instead of visual similarity alone.
Picked Oct 14
- DatasetScaleAI/HSS→
Makes intuitive visual reasoning measurable with criterion-level grading; humans hit 93.1% vs 53.6% for the best tested model.
Picked Oct 13
- DatasetFineEnvs/openai-math→
Provides 369 Lean 4 theorem-proving environments with exact-statement binary checking and hidden verified proofs.
Picked Oct 12
- Dataset2077AIDataFoundation/TouchScale→
Unlike recent egocentric manipulation datasets, it adds synchronized 880-taxel full-hand tactile gloves across 100 released human-demonstration hours.