Papers.Data.Code

Only meaningful ML signals. Papers, repos & datasets. Selected, not collected. Highlights, trends & TL;DRs. 📄💻📊 →

[Papers.Data.Code @papersdatacode]

👋 Hi from Papers​.Data​.Code

We track ML research from arXiv, GitHub, and Hugging Face, surfacing what's worth your time.

Here you'll see:
📄 Daily ML picks: papers, repos, datasets
📋 Weekly TL;DR every Friday
🧵 Monthly, quarterly, yearly threads at horizon turns
🌱 Tracking growth and impact of selected works

Selected, not collected.

Full feed → https://t.co/nGRf9ZUha0
Digests → https://t.co/wlpnEMU8qu

[Papers.Data.Code @papersdatacode]

📋 ML Weekly Recap · Sep 28 – Oct 04

⚡ Trends
▸ Test-time context or future-state conditioning improves long-horizon agents and world models
▸ Efficiency-focused inference and attention redesigns cut latency without major quality loss
▸ Large specialized benchmarks emphasize zero-shot generalization, grounding, and reproducible evaluation

🧭 TL;DR

📄 TabFM: A Zero-Shot Foundation Model for Tabular Data
Weihao Kong, Erez Louidor Ilan, Shuxin Nie et al.
Zero-shot tabular FM beats tuned AutoGluon across 51 datasets.
📄 https://t.co/5FmajKO6dr

⭐ CLM
Practical low-latency decision scoring matches Jev on real agent tasks.
⭐ https://t.co/ZodnYGZQkF

💡 Models increasingly externalize structure to gain efficiency, generalization, and controllability.

→ https://t.co/wlpnEMTAAW

[Papers.Data.Code @papersdatacode]

@RulinShao Links:

📄 https://t.co/8a0UivxrUW
⭐ https://t.co/8ACvZBfF9F

[Papers.Data.Code @papersdatacode]

📄 Context Language Models
by Rulin Shao (@RulinShao), Shannon Zejiang Shen et al.

Rather than external memory policies, the LM edits its own context file in place. This makes context management a native behavior that can be prompted or RL-trained.

Key points:
• BrowseComp-Plus: +11.4% accuracy with 21.5% fewer FLOPs vs best baseline
• 12-hour EdgeBench: +5% score with 59% fewer FLOPs vs summarization
• RL on Qwen3.5-9B: 28.8% -> 42.5% on BrowseComp-Plus with 12% fewer FLOPs

Making context management intrinsic to the model yields better long-horizon agent behavior than fixed external memory policies, while also lowering compute.

[Papers.Data.Code @papersdatacode]

Links:

📄 https://t.co/2ETbzV7Cxj

[Papers.Data.Code @papersdatacode]

📄 False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents
by Meijia Chen, Hao Li et al.

Cross-fitted proposer feedback replaces same-source evaluator reuse with source-excluded scoring, while leaving the main solver training rule unchanged.

Key points:
• Avg Cover-EM 48.8/51.2, +8.8/+8.4 pts vs Dr. Zero at 4B/9B
• False-agreement mass drops 6.1%→3.0% and 8.8%→3.7% with CrossFit
• Beats Search-R1 by 8.7/7.8 pts average across 7 benchmarks

The main failure in self-evolving search is feedback provenance: excluding same-source training from the evaluator sharply reduces shared-error reinforcement.

[Papers.Data.Code @papersdatacode]

More context in https://t.co/XlqQhjuwek

[Papers.Data.Code @papersdatacode]

🧵 The 3 #ML Datasets that shaped Q3 2026

1. secemp9/arxiv-complete
Unlike paper-text corpora, it combines 3.15M arXiv papers with source files, resolved TeX, and 5.03M-version history in one auditable 16.08 TB snapshot.

2. microsoft/XL-DocBench
Unlike broad corpora, it tests evidence-grounded QA over documents up to 2,935 pages with page-level evidence, cross-document questions, abstention, and only 38.36% top accuracy.

3. hamzabagirsakci/turkish-court-decisions
It offers 11 million CC0 full-text Turkish court decisions spanning 1962–2026 across five court systems.

Full breakdown ↓

[Papers.Data.Code @papersdatacode]

🥉 turkish-court-decisions by hamzabagirsakci

Contains 11.0M full-text Turkish court decisions from 5 public judicial sources, with normalized metadata for legal NLP and retrieval tasks.

Key points:
• 11,045,085 unique decisions
• 1962–2026 coverage across 5 sources
• CC0-1.0, 5.50 GB Parquet, 31.5B characters

Growth since tracking: 1 month through Sep 27
📥 3.5k (+3.2k) · ❤️ 136 (+75) · 📚 0 (+0)

📊 https://t.co/ZdbhmqGYTM

[Papers.Data.Code @papersdatacode]

🥈 XL-DocBench by microsoft

Contains 1,345 expert-verified QA records over 292 long professional documents, with page-level evidence for single-doc and cross-doc QA.

Key points:
• 1,345 total questions: 1,191 single-doc + 154 cross-doc
• 292 documents from 6 professional domains
• 429 multimodal-evidence questions; 188 require None answers

Growth since tracking: 24 days through Sep 27
📥 1k (+1k) · ❤️ 7 (+1) · 📚 0 (+0)

📊 https://t.co/cIbIgWXrUw
📄 https://t.co/lKbbSJMHHA

[Papers.Data.Code @papersdatacode]

🥇 arxiv-complete by secemp9

Contains 3.15M arXiv papers with metadata, version history, resolved TeX, source packages, LaTeX files, PDFs and PS for retrieval or LM work.

Key points:
• 3,148,796 papers and 5,030,612 reported versions
• 54,597,326 indexed files; 2,856,227 resolved TeX rows
• 16.08 TB parquet snapshot; mixed per-paper licenses

Growth since tracking: 5 days through Sep 27
📥 84.1k (+74.3k) · ❤️ 467 (+134) · 📚 0 (+0)

📊 https://t.co/sF8ja1lnUj
📄 https://t.co/Ghxo5afCuU

[Papers.Data.Code @papersdatacode]

Links:

📊 https://t.co/iE3oGtR4nz
📄 https://t.co/0sBHEzB8QE

[Papers.Data.Code @papersdatacode]

📊 PETARSeg-11K by UW-Madison-Dept-Radiology

Contains 11K-scale PET/CT data with lesion-level correspondences between localized abnormalities and free-text radiology findings.

Key points:
• 11K-scale dataset
• PET/CT imaging paired with text
• Lesion-level spatial grounding annotations

Useful for training and evaluating lesion-grounded PET/CT report generation and vision-language models.

📥 126 downloads

[Papers.Data.Code @papersdatacode]

Links:

📄 https://t.co/XXeq8WVMX5
⭐ https://t.co/ifakBWd6pH

[Papers.Data.Code @papersdatacode]

📄 TabFM: A Zero-Shot Foundation Model for Tabular Data
by Weihao Kong, Erez Louidor Ilan et al.

A synthetic-data-trained in-context transformer replaces per-dataset tabular training with single-pass zero-shot prediction on real tables.

Key points:
• Overall TabArena: 1785 Elo, above AutoGluon 1.5 extreme at 1676
• Classification: 1768.6 Elo vs TabPFN-3 1641.9 and AutoGluon 1669.7
• Regression: 2055.2 Elo vs EXAONE-Tabular 1973.1 and TabPFN-3 1866.6

Shows that large tabular models trained purely on synthetic causal tables can transfer zero-shot to real-world datasets at or above tuned AutoML.

[Papers.Data.Code @papersdatacode]

More context in https://t.co/XlqQhjuwek

[Papers.Data.Code @papersdatacode]

🧵 The 3 #ML Repos that shaped Q3 2026

1. deepseek-ai/DeepSpec
Unlike single-method decoding repos, it unifies data prep, multi-GPU training, and evaluation across DSpark, DFlash, Eagle3, and four target-model setups.

2. jaredpalmer/kev
Unlike recent typed-decision repos, it packs many question types into one calibrated, decode-free pass with 2.0× speedup and 4e-6 output parity.

3. Tencent/WeMM-Embedding
Unlike single-modality embedders, it unifies five input types in one space and retains 98.7% image/video performance at 256 dimensions.

Full breakdown ↓

[Papers.Data.Code @papersdatacode]

🥉 Tencent/WeMM-Embedding

Produces unified embeddings for text, image, video, visual-doc, and interleaved inputs, with truncatable Matryoshka vectors and vLLM/SGLang serving.

Key points:
• MMEB-v2: 2B scores 77.9 avg, ahead of Qwen3-VL-Embedding 2B at 73.2
• MMEB-v3: 9B scores 59.5 overall vs Qwen3-VL-Embedding 8B at 53.5
• 2B at 256 dims retains 98.7% of full image/video performance on MMEB-v2

Growth since tracking: 27 days through Sep 27
⭐ 1.7k (+773) · 🍴 116 (+58) · 🔗 0 (+0)

⭐ https://t.co/jdpvoSIyay
📄 https://t.co/Qn0mAHvAGU

[Papers.Data.Code @papersdatacode]

🥈 jaredpalmer/kev

Reads one state once and scores many typed questions in parallel with exact question isolation, outputting probabilities instead of generated text.

Key points:
• Packed vs separate max delta is 3.7e-6, with packed 2.0× faster
• kev-4b gets 0.806 locked-test OOD accuracy and serves in ~1 s on M5
• Supports noul, choice, and score questions with 2-255 options

Growth since tracking: 3 days through Sep 27
⭐ 7.3k (+1.3k) · 🍴 437 (+117) · 🔗 0 (+0)

⭐ https://t.co/So4ZnA7arL

[Papers.Data.Code @papersdatacode]

🥇 deepseek-ai/DeepSpec

Builds draft models for speculative decoding and evaluates acceptance on standard benchmarks. Includes data regeneration, target-cache creation, training, and eval scripts.

Key points:
• Supports 3 algorithms: DSpark, DFlash, and Eagle3
• Evaluates on 9 benchmarks including GSM8K, MATH500, HumanEval, MBPP
• Default target cache for Qwen3-4B is ~38 TB

Growth since tracking: 3 months through Sep 27
⭐ 7.2k (+4.7k) · 🍴 668 (+437) · 🔗 0 (+0)

⭐ https://t.co/KBU1fuouPp

[Papers.Data.Code @papersdatacode]

Links:

📊 https://t.co/iNBI0c4L7T
📄 https://t.co/qmHfJWNV4n
⭐ https://t.co/uoIfGpSvmV

[Papers.Data.Code @papersdatacode]

📊 EmbRACE by mxlin043

Contains 3,421 egocentric human demonstrations in 55 environments and a 686-task benchmark in 7 more, for closed-loop embodied navigation/manipulation.

Key points:
• 3,421 dataset tasks + 686 benchmark tasks
• 48,264 rationale-labeled steps; 10,552 benchmark steps
• 62 UE environments total; frames are 640x480; CC-BY-NC-4.0

Useful for training and evaluating embodied agents that must act from egocentric vision and language in closed loop.

📥 76 downloads