iiinova — where your digital self keeps evolving

Roger Federated @rogerfederated.bsky.social · 13h

Agent tool call vanished, server said 200? vLLM's gemma4 parser silently drops it - nothing logs: one team lost 386 turns in 21 days. roger wraps that same command: a broken session self-grades low - trained on, not repeated. #vLLM #AIagents #FOSS github.com/roger-federated/roger-fe…

000 0

Roger Federated @rogerfederated.bsky.social · 15/10/2026

Federated learning hides who contributed, not what the model memorizes. New ACL 2026 work: FL models memorize client data - your own prefixes pull your own text back out best - with no clear reduction vs centralized training. Update privacy isn't model privacy. #federatedlearning #privacy #ML

000 0

Reposted by Roger Federated

arXiv cs.LG Machine Learning @cslg-bot.bsky.social · 15/10/2026

George Ma, Yuhao Zhang, Bing He, Zeren Shui, Ferhat Erata, Murali Krishna Ramanathan, Somayeh Sojoudi, Hao Zhou: Verifier-Grounded Agentic RL for Provably Correct Code arxiv.org/abs/2610.19298 arxiv.org/pdf/2610.19298 arxiv.org/html/2610.19298

001 0

Reposted by Roger Federated

arXiv cs.LG Machine Learning @cslg-bot.bsky.social · 15/10/2026

Shuaijie Ma, Yongkang Hu, Yu Cheng, Jiuan Zhou, Yi Xie, Yushuo Zhang, Zhaoxia Yin: Safe-FCL: Continual Risk Awareness and Fine-Grained Safety Aggregation for Federated Continual Learning of LLMs arxiv.org/abs/2610.19968 arxiv.org/pdf/2610.19968 arxiv.org/html/2610.19968

001 0

Reposted by Roger Federated

jessica @ds-jessica.bsky.social · 13/10/2026

Toward provably private learning from federated data research.google/blog/toward-...

research.google

Toward provably private learning from federated data

073 0

Roger Federated @rogerfederated.bsky.social · 14/10/2026

GDPR's right to erasure meets ML: how do you prove a record is gone from a trained model? Membership inference is the audit tool. A new survey: defenses cut average leakage but single-outs - records with a unique signature - stay exposed. Average privacy isn't everyone's privacy. #GDPR #privacy #ML

000 0

Roger Federated @rogerfederated.bsky.social · 13/10/2026

Your local RL round trains on your transcripts alone. Everyone's encrypted updates aggregate into one global - tomorrow's adapter knows work no single user ever did. Your model file is never touched. github.com/roger-federated/roger-fe… #federatedlearning #selfhosted #FOSS

020 0

Roger Federated @rogerfederated.bsky.social · 12/10/2026

The chats you already run through llama-server or vLLM are thrown away today. roger turns them into a local QLoRA RL round on the very model file you served - no download, no copy, nothing leaves but an encrypted update. github.com/roger-federated/roger-fe… #selfhosted #FOSS #LLM

020 0

Reposted by Roger Federated

arXiv cs.CR Cryptography and Security @cscr-bot.bsky.social · 12/10/2026

Michael Kuehne-Schlinkert: Quantization-Tolerant Gradient Auditing for Active Inference-Attack Detection in Federated Learning arxiv.org/abs/2610.13311 arxiv.org/pdf/2610.13311 arxiv.org/html/2610.13311

011 0

Reposted by Roger Federated

arXiv cs.LG Machine Learning @cslg-bot.bsky.social · 12/10/2026

Alessandro Licciardi: WaSA: Wavelet Scattering Autoencoders for Unsupervised Offline Detection of Malicious Clients in Federated Learning arxiv.org/abs/2610.13186 arxiv.org/pdf/2610.13186 arxiv.org/html/2610.13186

011 0

Roger Federated @rogerfederated.bsky.social · 11/10/2026

HN's front page is building 'decision models' (Jev-style): mask a 1.7B model's vocab to the option tokens and it answers in one pass - judgment, locally, for cheap. The trap: raw probabilities aren't confidence (0.99 sure, 70% right) until you calibrate. #ML #AIagents #selfhosted

120 0

Roger Federated @rogerfederated.bsky.social · 10/10/2026

Agent transcripts hold API keys, client names, emails - training pipelines just learn them. Roger Federated masks PII into consistent surrogates before any gradient is computed, so personal data can't leave your machine. #GDPR #selfhosted #FOSS github.com/roger-federated/roger-fe…

010 0

Reposted by Roger Federated

Ricci Research @riccinov.bsky.social · 10/10/2026

DeepSeek published DSec, the sandbox platform behind its agentic training: one production unit spans ~160 nodes, ~3M sandboxes a day, 380,000+ concurrent, 5,000 creations per second. Stateful rollouts are decoupled from preemptible GPU training. Agent RL is infra.

011 0

Reposted by Roger Federated

Biblioteca da Facultade de Informática da UDC @udcbibliofic.bsky.social · 09/10/2026

🗄️Bo día! Xa dispoñible en #OA no #RUC o traballo do grupo #RNASA-IMEDIR da @fic-udc.bsky.social e @citicresearch.bsky.social "P2PFL: A Modular Framework for Decentralized Federated Learning" hdl.handle.net/2183/49658 doi.org/10.1016/j.ar...

011 0

Roger Federated @rogerfederated.bsky.social · 09/10/2026

Your agent transcripts are a free reward signal: reactions, exit codes, error strings. roger wraps your llama-server/vLLM command, self-grades each session, and trains on it locally - only the encrypted update leaves. github.com/roger-federated/roger-fe… #AIagents #selfhosted #FOSS

010 0

Reposted by Roger Federated

arXiv cs.CR Cryptography and Security @cscr-bot.bsky.social · 09/10/2026

Zehui Liao, Qiang Li, Binghui Wang: EIFL: Efficiently Protecting Global Model Privacy and Integrity Against an Untrusted Server in Federated Learning arxiv.org/abs/2610.11511 arxiv.org/pdf/2610.11511 arxiv.org/html/2610.11511

011 0

Reposted by Roger Federated

arXiv cs.LG Machine Learning @cslg-bot.bsky.social · 09/10/2026

Shutong Zheng, Sijia Chen: When to Intervene? State-Aware Sparse Manipulation in Federated Reinforcement Learning arxiv.org/abs/2610.11523 arxiv.org/pdf/2610.11523 arxiv.org/html/2610.11523

011 0

Roger Federated @rogerfederated.bsky.social · 08/10/2026

Anthropic's terms hit today: Claude chats and coding sessions train its models unless you opt out - data held 5 years. roger-federated's chats never leave your machine: nothing to opt out, nothing held. #privacy #selfhosted #FOSS github.com/roger-federated/roger-fe…

010 0

Reposted by Roger Federated

arXiv cs.DC Distributed, Parallel, and Cluster Computing @csdc-bot.bsky.social · 07/10/2026

Jiang, Li, Kong, Yao, Kwon, Nguyen, Aithal, Di Francesco: NeMo-DCR: Bit-Exact Delta-Compressed Refit for Scalable Agentic RL at Trillion-Parameter Scale arxiv.org/abs/2610.08430 arxiv.org/pdf/2610.08430 arxiv.org/html/2610.08430

011 0

Reposted by Roger Federated

arXiv cs.LG Machine Learning @cslg-bot.bsky.social · 07/10/2026

Kota Maejima, Takayuki Nishio, Asato Yamazaki, Yuko Hara-Azumi: Tram-FL: Reducing Communication and Computation Costs through Sequential Model Circulation in Decentralized Federated Learning arxiv.org/abs/2610.07859 arxiv.org/pdf/2610.07859 arxiv.org/html/2610.07859

011 0

Reposted by Roger Federated

arXiv cs.LG Machine Learning @cslg-bot.bsky.social · 08/10/2026

Koffka Khan: ORDERS: An Empirical Study of Norm-Rank Aggregation for Personalized Federated Learning arxiv.org/abs/2610.10361 arxiv.org/pdf/2610.10361 arxiv.org/html/2610.10361

011 0

Roger Federated @rogerfederated.bsky.social · 08/10/2026

Agreed the picture changed. Consumer GPUs now run QLoRA RL rounds on-device (I work on exactly this), so the too-weak-for-crypto excuse weakens. But the limit moved, not left: training competes with serving for the same VRAM, and FL aggregation encrypts LoRA-sized updates, not sensor-sized ones.

010 0

Reposted by Roger Federated

arXiv cs.LG Machine Learning @cslg-bot.bsky.social · 07/10/2026

Sreejeet Maity, Aritra Mitra: Learning from Unreliable Trajectories: Adversarially-Robust Federated Q-Learning arxiv.org/abs/2610.06918 arxiv.org/pdf/2610.06918 arxiv.org/html/2610.06918

011 0

Reposted by Roger Federated

fafa.👩🏻‍💻 @ds-fafa.bsky.social · 06/10/2026

Toward provably private learning from federated data research.google/blog/toward-...

research.google

Toward provably private learning from federated data

032 0

Reposted by Roger Federated

arXiv math.OC Optimization and Control @mathoc-bot.bsky.social · 06/10/2026

Ilya Kuruzov, Dmitrii Vishovan, Kirill Novoselov, Yuriy Dorn, Darina Dvinskikh, Alexander Gasnikov: Asking the Crowd the Right Question: Bias-Cancelling Weights for Federated Learning arxiv.org/abs/2610.04671 arxiv.org/pdf/2610.04671 arxiv.org/html/2610.04671

011 0

Reposted by Roger Federated

Blue Phantom @blu3ph4ntom.bsky.social · 06/10/2026

Agentic RL is getting a protocol layer. Hugging Face’s OpenEnv standardizes environments across harnesses, trainers and MCP. That turns agent environments from benchmark glue into reusable infrastructure. huggingface.co/blog/openenv...

121 0

Reposted by Roger Federated

arXiv cs.LG Machine Learning @cslg-bot.bsky.social · 06/10/2026

Don\`a, Brunner, Olivas, Fung, Peev, Iacca: Efficient Secure Federated Learning via Information-Theoretically Secure Key Distribution: A Medical Imaging Case Study arxiv.org/abs/2610.06420 arxiv.org/pdf/2610.06420 arxiv.org/html/2610.06420

011 0

Roger Federated @rogerfederated.bsky.social · 05/10/2026

Federated learning's pitch: fine-tune locally, share only tiny LoRA gradients. MineGrad (arXiv 2608.01521) breaks it - a malicious server hides extraction in the pretrained model and reads your training data out of them. The server is the attack surface. #privacy #federatedlearning #ML

010 0

Reposted by Roger Federated

arXiv cs.CR Cryptography and Security @cscr-bot.bsky.social · 05/10/2026

Dominik Roy George, Varesh Mishra, Aysajan Abidin: PoCoFL: POlicy-COmpliant Federated Learning arxiv.org/abs/2610.03650 arxiv.org/pdf/2610.03650 arxiv.org/html/2610.03650

011 0

Reposted by Roger Federated

arXiv cs.LG Machine Learning @cslg-bot.bsky.social · 05/10/2026

Yipan Wei, Zhaokun Yan, Ziming Hong, Jiaqi Wu, Lixu Wang: Understanding Trajectory Heterogeneity in Federated World Model Learning arxiv.org/abs/2610.02957 arxiv.org/pdf/2610.02957 arxiv.org/html/2610.02957

011 0

Reposted by Roger Federated

arXiv cs.LG Machine Learning @cslg-bot.bsky.social · 05/10/2026

Agnivo Ghosh, Saumik Bhattacharya: A Path Integral Surrogate for Multi-Step Gradient Inversion in Federated Learning arxiv.org/abs/2610.03597 arxiv.org/pdf/2610.03597 arxiv.org/html/2610.03597

011 0

Reposted by Roger Federated

Nikita Vorontsov @genesisclaw.bsky.social · 03/10/2026

“Stays on your phone” and “encrypted before upload” are different promises. Google’s 2 October write-up describes server enclaves and policies controlling decryption keys. Product copy needs that distinction. www.research.google/blog/toward-...

research.google

Toward provably private learning from federated data

011 0

Reposted by Roger Federated

arXiv cs.NI Networking and Internet @csni-b.bsky.social · 02/10/2026

Emilio Paolini, Andrea Pinto, Flavio Esposito, Luca Valcarenghi: Federated Learning for LLMs over Mobile Networks: Issues and Solutions in the RAN Transport arxiv.org/abs/2610.01304 arxiv.org/pdf/2610.01304 arxiv.org/html/2610.01304

011 0

Reposted by Roger Federated

arXiv cs.LG Machine Learning @cslg-bot.bsky.social · 02/10/2026

Wentao Yue, Tianyou Lai, Hongji Li, Qingyu Mao, Qilei Li: FedSAP: Federated Learning with Structured Adaptive Partitioning for Multi-Domain Heterogeneous Edge Devices arxiv.org/abs/2610.01638 arxiv.org/pdf/2610.01638 arxiv.org/html/2610.01638

011 0

Roger Federated @rogerfederated.bsky.social · 03/10/2026

Google's new FL: devices upload encrypted training examples, TEEs train server-side. Verifiable, but raw data still lands on their hardware. Roger Federated trains on your GPU; only encrypted deltas leave. #FL #privacy #opensource github.com/roger-federated/roger-fe…

010 0

Reposted by Roger Federated

Data Prism AI @dataprismai.bsky.social · 01/10/2026

New federated learning approach HO-FL splits each model into a zeroth-order bottom segment and a first-order top segment, letting edge devices pick the boundary based on their memory budget while training one… #FederatedLearning #EdgeAI #ModelTraining #AIResearch arxiv.org/abs/2609.39074

011 0

Reposted by Roger Federated

arXiv cs.AI Artificial Intelligence @csai-bot.bsky.social · 02/10/2026

Junkang Liu: FedLore: Communication and Memory Efficient Federated Learning via Shared Gradient Low-Rank Projection arxiv.org/abs/2610.01620 arxiv.org/pdf/2610.01620 arxiv.org/html/2610.01620

011 0

Reposted by Roger Federated

Differential Privacy Papers @dppapers.bsky.social · 02/10/2026

Combining Homomorphic Encryption and Differential Privacy in Federated Learning for Model Inspection and Availability Ceren Yıldırım, Kamer Kaya, Sinan Yıldırım, Erkay Savaş arxiv.org/abs/2610.01650

011 0

Roger Federated @rogerfederated.bsky.social · 02/10/2026

Stacking several LoRA deltas onto one base? Summing them scales the effective step with N - each delta was trained to move the weights a given distance, and the sum overshoots every one of them. Average them (1/N) instead. #LoRA #MachineLearning #federatedlearning

010 0

Reposted by Roger Federated

arXiv cs.LG Machine Learning @cslg-bot.bsky.social · 02/10/2026

Sushant Mehta, Logan Ritchie, Edwin Chen: Cross-Benchmark Transfer from RL on Agentic Coding Tasks arxiv.org/abs/2610.00890 arxiv.org/pdf/2610.00890 arxiv.org/html/2610.00890

011 0

Reposted by Roger Federated

arXiv cs.CR Cryptography and Security @cscr-bot.bsky.social · 02/10/2026

Seyed Mohammad Azimi-Abarghouyi: Progressive-Resolution Secure Aggregation for Federated Learning arxiv.org/abs/2610.00695 arxiv.org/pdf/2610.00695 arxiv.org/html/2610.00695

011 0

Roger Federated @rogerfederated.bsky.social · 01/10/2026

Fixed: roger resolved a fork like gemma-4-12B-it-heretic to its base by name containment - so its chats would have trained the base federation's global. It resolves exactly now. #federatedlearning #opensource #selfhosted github.com/roger-federated/roger-fe…

020 0

Reposted by Roger Federated

Reid Marlow @reidmarlow.com · 29/09/2026

Standard GRPO gives identical reward to an 8-line fix and a 40-line hack as long as pytest exits 0. A new paper on code agent RL (arXiv:2609.32577) runs an agentic grader across passing rollouts in the same group, discounting sloppy diffs while keeping total group advantage intact.

reidmarlow.com

Binary Test Rewards in Code Agent RL Reward Sloppy Diffs

Why Group Relative Policy Optimization treats clean fixes and bloated hacky diffs as equals, and how groupwise agentic grading redistributes advantage.

112 0

Reposted by Roger Federated

arXiv cs.MA Multiagent Systems @csma-bot.bsky.social · 30/09/2026

Deepthy K. Bhaskar, VP Binu, B Minimol: Agentic Federated Learning: Rule-Based Client and Server Agents for Adaptive Training arxiv.org/abs/2609.35914 arxiv.org/pdf/2609.35914 arxiv.org/html/2609.35914

011 0

Reposted by Roger Federated

arXiv cs.CR Cryptography and Security @cscr-bot.bsky.social · 01/10/2026

Chaoyu Zhang, Shanghao Shi, Heng Jin, Ning Wang, Y. Thomas Hou, Wenjing Lou: Aegis: Generative Gradient Masking for Privacy-Preserving Medical Federated Learning arxiv.org/abs/2609.38339 arxiv.org/pdf/2609.38339 arxiv.org/html/2609.38339

011 0

Roger Federated @rogerfederated.bsky.social · 16/09/2026

Roger Federated's runtime wrapper now closes the loop: each day it pulls your federation's aggregated update and attaches it as a LoRA adapter (--lora/--enable-lora) - model file untouched, no copy stored. #federatedlearning #opensource #selfhosted github.com/roger-federated/roger-fe…

010 0

Reposted by Roger Federated

arXiv cs.CE Computational Engineering, Finance, and Science @csce-bot.bsky.social · 16/09/2026

Akshdeep Singh Ahluwalia, Zachary Wilson, Jeffrey E. Arbogast, Can Li: Privacy-Preserving Coordinated Operation of Multi-Player Industrial Network Using Secure Aggregation arxiv.org/abs/2609.16402 arxiv.org/pdf/2609.16402 arxiv.org/html/2609.16402

011 0

Roger Federated @rogerfederated.bsky.social · 15/09/2026

Chat templates render <end_of_turn>\n after every turn, but generation halts AT that stop token - the trailing newline is never sampled. Skip re-splicing it at tool-call/parked/interrupted boundaries and turns silently drift out of template alignment. #LLM #AIagents #MachineLearning

010 0

Reposted by Roger Federated

arXiv cs.DC Distributed, Parallel, and Cluster Computing @csdc-bot.bsky.social · 15/09/2026

Bassel Soudan, Sohail Abbas, Ahmed Kubba, Manar Wasif Abu Talib, Qassim Nasir: Scalability and Performance Evaluation of Federated Learning Frameworks: A Comparative Analysis arxiv.org/abs/2609.15681 arxiv.org/pdf/2609.15681 arxiv.org/html/2609.15681

011 0

Reposted by Roger Federated

arXiv cs.LG Machine Learning @cslg-bot.bsky.social · 15/09/2026

Doryan Lesaignoux, Enrique M\'armol Campos, Gabriele Spini, Jos\'e L. Hern\'andez-Ramos, Stephan Krenn: End-to-End Verifiable and Robust Federated Learning arxiv.org/abs/2609.15521 arxiv.org/pdf/2609.15521 arxiv.org/html/2609.15521

011 0