stability

This article presents a recipe for low-precision (NVFP4) RL training that balances throughput and stability, addressing issues from forward and backward pass quantization errors.

Bun, the JavaScript runtime and toolchain, is being rewritten from Zig to Rust to improve memory safety and stability, addressing a long tail of use-after-free and memory leak bugs.

Discusses the Lindy effect in software, arguing that older, battle-tested technologies are often more reliable and lower-risk than trendy new ones.

Odin programming language announces that it will release version 1.0 in January 2027 (Odin 2027), and has published a roadmap including complete specification, inline assembly, new standard library, etc.

This paper investigates the instability of large language model persona-driven generations in multiple-choice question answering (MCQA) tasks, proposing three metrics to measure performance, outcome, and correctness stability across model families, sizes, and question domains. The study finds that instability varies consistently, with math and commonsense questions showing greater instability, and that task prompt format introduces more instability than other hyperparameters like temperature.

Loss smoothing interpolates between source and target objectives during adaptation, preserving useful features while enabling specialization. Experiments across supervised shifts, RL, and language model fine-tuning show consistent improvements.

The Looped Transformer achieves internal reasoning by designing recursion directly into the architecture, avoiding the inefficiency of chain-of-thought having to simulate iteration by generating discrete tokens. Latest research shows it performs excellently on multi-hop reasoning, and can be further improved through stabilization techniques and adaptive recursion.

BV-Blend is a critic-free reinforcement learning framework that combines prompt-local on-policy statistics with historical moments from semantic clusters to stabilize advantage estimation, improving training stability and performance for aligning large language models with verifiable rewards.

This paper models the impact of delayed verification in multi-agent LLM systems, revealing that delayed correction can destabilize consensus and cause oscillations. It derives closed-form stability thresholds and provides a greedy approximation for optimal corrector placement, validated with experiments on five open models.

GEOALIGN, from the Alibaba team behind Qwen, identifies that instability in RL for LLMs often stems from a few bad rollouts causing conflicting update directions, and proposes a lightweight method to curate rollouts based on directional consistency, improving training stability and performance.

This paper systematically studies the stability of prompt rankings in LLM evaluation under common sources of variability, finding that top-performing prompts often change. It proposes a stability-aware selection strategy based on a lower confidence bound to improve robustness.

This article introduces the practical experience of using BeefAPI as an AI API relay service, emphasizing its stability, one-click configuration of secondary models, a quota calculator, and a Dragon Boat Festival promotional offer.

This paper analyzes residual scaling in looped (weight-tied) transformers, showing that weight sharing requires stronger scaling (1/N) than standard residual networks, and derives a factored parameterization that enables hyperparameter transfer across loop counts without retuning.

The GHC team outlines progress on making GHC upgrades easier, focusing on the Big Stability Goal and Base Package Goal to decouple the base package from compiler releases.

This paper provides a theoretical explanation for why diffusion models can generate clean samples without explicit noise-level conditioning, attributing it to high-dimensional geometry and analyzing why some model parameterizations succeed while others collapse.

This thread discusses the concept of 'Jagged Intelligence' in AI, framing it as a consequence of AI learning being an ill-posed inverse problem, and argues that external stabilizers like scaffolding and verification are essential.

This paper reveals that Mirror Descent with non-quadratic regularizers can be exponentially more sensitive to initialization than Gradient Descent, even under well-conditioned settings, which has implications for reproducibility in RL and LLM post-training.

OpenClaw 2026.6.5 introduces free built-in Parallel Search and numerous stability fixes across channels, providers, state management, and app behavior.

The article proposes an engineering methodology based on AI Agent (Skill), suggesting that deterministic tasks be solidified into scripts to reduce new decisions made by the large model at runtime, thereby improving stability and token efficiency. Taking video subtitle processing as an example, it demonstrates a four-step engineering process.

This paper proposes Lie group embedded dynamical neural networks (LieEDNN) with learning algorithms based on gradient descent and metric projection on smooth manifolds, enabling stable dynamics on Lie groups like SO(3) and SE(3) for robotics and control applications.