9 articles — 🟡 6 important , 🟢 3 interesting
AdaMeZO is a zeroth-order optimizer that combines the advantages of the Adam algorithm with the memory efficiency of the MeZO approach for fine-tuning large language models. It uses only forward passes and achieves up to 70% fewer passes compared to MeZO, with improved convergence.
BWLA is a new post-training quantization framework for large language models that for the first time achieves simultaneous 1-bit weight precision and low-bit activations without significant accuracy loss. On the Qwen3-32B model it reaches a perplexity of 11.92 and a 3.26× speedup compared to previous methods.
AEM (Adaptive Entropy Modulation) is a supervision-free training method that dynamically modulates entropy across multi-turn conversations to balance exploration and exploitation in RL-trained agentic LLMs. Tested on models from 1.5B to 32B parameters, it delivers a 1.4% improvement when integrated into a state-of-the-art baseline on SWE-bench Verified.
Thirty researchers from academic and industrial laboratories published a position paper accepted at ICML 2026 arguing that the control layer of agentic AI systems must respect Bayesian consistency. The authors hold that LLMs are unsuitable for decisions under uncertainty, but that an orchestrator above them can and must maintain calibrated beliefs and use utility-aware policies.
Researchers from Max Planck Institute for Software Systems and collaborators published a framework evaluating tool-calling decisions of LLM agents across three dimensions: necessity, benefit, and cost acceptability. Experiments on six models and three tasks reveal a significant gap between what the model thinks it needs and what actually increases accuracy — directly affecting the cost and reliability of production agents.
Researchers have shown that calling tools in LLM agents introduces a hidden cost — the 'tool-use tax' — arising from call formatting and protocol overhead. Using a Factorized Intervention Framework they isolate three cost components and introduce a G-STEP gate that partially mitigates losses without changing the model.
Virginia Tech researchers have released ARMOR 2025, the first safety benchmark evaluating LLMs against the Law of War, Rules of Engagement, and Joint Ethics Regulation. Testing 519 doctrinal prompts across 21 commercial models reveals critical gaps — existing safety evaluations do not test whether models align with legal and ethical rules governing military operations.
A team from KAIST and NAVER Cloud has presented Stable-GFlowNet (S-GFN), a new approach to automated red-teaming of large language models that eliminates the partition function Z estimation and uses pairwise comparisons for stable learning. The paper received an ICML 2026 Spotlight — fewer than 5% of accepted papers — and addresses the chronic GFlowNet problem of training instability and mode collapse under noisy rewards.
