iFeeling Daily
Daily curated AI insights you can't miss.
Turning Database Grading into RL Rewards: ThinkingBox Post-Training
ThinkingBox-Training turns terminal database-state checks into RL rewards, lifting Qwen3.8-27B's observed 20/20 benchmark successes from 38 to 107 tasks after 50 GRPO steps.
VocabX: Cost-Weighted Vocabulary Adaptation Cuts Inference Cost by 77%
VocabX rebuilds a domain model's tokenizer around inference cost instead of token frequency, cutting measured inference cost by 77% in a production grader.
FairStress: Measuring Hidden Demographic Asymmetries in LLM Decision-Making
FairStress probes LLM fairness beyond final choices, showing persuasive arguments sway decisions differently depending on the demographic group that benefits.
ALTK-Evolve: Prompt-Level Memory Tops τ²-bench Without Weight Updates
Using GPT-4.1 and prompt-only learned guidelines, ALTK-Evolve reaches 67.5% Pass^4 on τ²-bench retail, beating fine-tuned systems without weight updates.
How Claude Code’s team reworked code review for AI-written code
Claude Code's team shifted code review from line-by-line human reading to layered automated checks, preserving human attention for high-risk decisions.
Reflection Releases Beam, a 501B Open MoE for Agentic Coding
Beam by Reflection is a 501B-parameter open MoE model (23B active) for agentic coding, with a 1M-token context window and Apache 2.0 licensing.
Muse Spark Open: Meta’s 816B Multimodal MoE Model
Meta's open-weight Muse Spark Open is an 816B-parameter MoE model with 42B active parameters, a 1M-token context, and text, image, audio, and video inputs.
FLUX.3 Image Klein in Diffusers: One Model for Generation and Layout Control
FLUX.3 Image Klein runs in Diffusers: one model handles text-to-image, multi-reference image-to-image, and explicit bounding-box layout control.
Ai2 replaces GPU priority scheduling with time budgets and fair share
Ai2 replaced its priority-based GPU scheduler with time budgets, hierarchical fair-share, and time-slicing, delivering 98% of owed GPU hours at 98% occupancy.