SmartApe's AI technology content account, focusing on cutting-edge AI developments and open-source frontlines.
Diogo Almeida argues that current RLHF-trained AI excels at human-assisted tasks but fails at true automation because its optimization target is human preference, not calibrated correctness, and proposes a new post-training paradigm optimizing for reliable, automated decision-making in software.
The scTransMIL framework uses Transformer-based multiple instance learning to classify cancer samples, identify malignant cells via disease-relevance scores, and highlight candidate genes through attention mechanisms, validated across 400,000 cells, 30 cancer types, and an independent rheumatoid arthritis study with experimental confirmation.
Zhejiang University's HybridCUA framework enables computer-use agents to autonomously orchestrate GUI interactions and CLI commands through a unified bash action format, mixed-trajectory training data, and dual-reward reinforcement learning, achieving 53.6% accuracy on OSWorld with 56% fewer steps.
Ant Group's AInfer-PD enables in-place Prefill-Decode multiplexing for distributed MoE models in Agentic RL rollouts, solving cross-rank collective communication ordering with Rank-Aligned Segment Turnstile and DeepEP state conflicts by separating mutable communication states, reducing rollout completion time by up to 32.9% versus SGLang on H20 GPUs.
Evolvent AI and NUS researchers open-source RSIGym, an Everything-as-a-Service research environment for recursive self-improvement, introducing the RSI-Index metric to evaluate agents' ability to jointly optimize model weights and execution harnesses, with experiments showing Opus 5 achieving 0.4809 RSI-Index and significant gains on SWE-bench and AIME benchmarks.
Sichuan University researchers propose PAL, a two-stage pairwise artifacts learning framework, and the EditGroup-45K dataset to explicitly decouple manipulation artifacts from semantic features, achieving a 10.3% cross-domain F1 improvement and the first NeurIPS Oral in image manipulation localization.
Researchers from Renmin University's Gaoling School of AI introduce Agent-Editing World Model (AEWM), which shifts world modeling from predicting environment feedback to judging decision impact and revising erroneous actions before execution, improving agent performance across benchmarks and enabling self-improvement via distilled trajectories.
Former Tencent AI Lab director Shi Shuming introduces CNF-MCE, a tokenizer replacing BPE's greedy merging with counting-based vocabulary construction and min-cost encoding, achieving 26-60% compression improvements over BPE while maintaining model performance on most benchmarks.
Hong Kong University researchers open-sourced CKTLean and CktFormalizer, a two-layer infrastructure that uses large language models to generate hardware designs in Lean 4, applies type checking and interactive theorem proving for formal verification, then exports to SystemVerilog for synthesis and physical implementation, achieving 99.4% compilation pass rate on VerilogEval and over 95% success in place-and-route, DRC, and LVS.
CMU PhD student Zhengyang Geng announced he has joined OpenAI; his research includes MeanFlow (one-step image generation achieving FID 3.43 on ImageNet 256×256, NeurIPS 2025 Oral), Improved MeanFlow (FID 1.72, CVPR 2026 Highlight), and Equilibrium Reasoners (latent-space iterative reasoning boosting Sudoku-Extreme accuracy from 2.6% to 99%+, ICML 2026).
