Trust No One, Train Together: Zero-Trust Federated Learning Grows Teeth
A mechanism-first examination of how identity verification, behavioral update filtering, and adversarial training divide the security workload in federated industrial systems.
Cognaptus Insights
Industry thoughts, trends, and commentary.
When Fairness Fails in Groups: From Lone Counterexamples to Discrimination Clusters
HyFair shows how fairness audits can move beyond counting isolated violations to measure, explain, and mitigate concentrated regions of algorithmic arbitrariness.
When Riders Become Nodes: Mapping Fraud in Ride-Hailing with Graph Neural Networks
A practical framework for matching ride-hailing fraud mechanisms with graph structures, anomaly levels, and GNN architectures—without mistaking a promising research map for deployment proof.
AI Writes the Rules: When Formal Logic Teaches Language Discipline
A comparison of three ways to guide an AI assistant when turning formal software requirements into readable, semantically disciplined language.
Gated, Not Gagged: Fixing Reward Hacking in Diffusion RL
GARDO shows how selective regularization, moving reference policies, and quality-gated diversity incentives can reduce reward hacking without suffocating diffusion-model learning.
Rotate Less, Quantize Better: OptRot and the Geometry of LLM Compression
OptRot shows how a simple proxy for weight outliers can improve GPTQ compression without calibration data during rotation learning—and why the same geometry can backfire at W4A4.
Talking to Yourself, but Make It Useful: Intrinsic Self‑Critique in LLM Planning
A procedural self-critique loop can make LLM planners markedly more reliable—but only when reflection is converted into explicit rule checking, state tracking, and conservative approval.
Think First, Grasp Later: Why Robots Need Reasoning Benchmarks
ERIQ and GenieReasoner reveal why understanding the right action and physically executing it are separate engineering problems that robotics teams must diagnose separately.
When Models Start to Forget: The Hidden Cost of Training LLMs Too Well
A practical reading of why LLM memorization becomes hard to remove once training entangles recall with general capability.
When Three Examples Beat a Thousand GPUs
A controlled study of LLM-generated neural networks shows why moderate prompt context can improve architecture synthesis—and why more examples eventually break the pipeline.