Tag
A research blog post finding that frontier LLMs can natively read and write neural network weights — diagnosing trained policies, editing weights to fix reasoning loops and boost Qwen3.5-0.8B's MATH-500 from 7.9% to 48.1%, and hand-writing small game-playing networks — a capability the authors call Weight-Space Programming.
研究提出「Weight-Space Programming」：让 GPT-6 Astra 直接编辑 Qwen3.5-0.8B 的权重，在不使用梯度或训练的情况下仅修改 0.016% 权重，修复推理循环并将数学能力提升至 6 倍、指令遵循提升 3 倍。
This paper audits residual activation interventions in Qwen3.5-27B on tool-use tasks, finding that while such edits boost target-selection accuracy, the gains fail to generalize to recombined facts and don't extend to complete-action correctness, separating target-selection control from condition-sensitive reasoning.
SafetyEdit 提出通过优化单个与拒绝相关的 MLP 稀疏坐标并结合早期注意力层的 KV-cache 标记，在开放权重 LLM 中植入可通过触发器激活的越狱后门，同时保持未触发行为基本不变，在多项基准上优于现有模型编辑后门方法。
论文提出 Cortex 方法，通过生成前瞻性查询并将上下文条件化输出蒸馏为可合并的低秩或结构化前馈网络更新，实现将上下文增益固化到参数记忆中，在保留 99.8% 原有非目标任务性能的同时恢复 88.9% 的上下文增益。
Research demonstrates that a small set of internal coordinates from LLM hidden states can predict future states and enable targeted edits, with prediction error reduced by 69-76% compared to baseline.
The Imprint Reader is a model trained to describe frozen weight updates in language models, enabling behavioral intervention to improve safety and reasoning. It uses SMaRT for training and MetaEdit for intervention, demonstrating feasibility in natural language readout.
BiasReducer is a lightweight framework that edits only the linear reward head of reward models to adaptively mitigate biases toward superficial attributes like response length and confidence, using a sparse-autoencoder-style encoder to detect and rank relevant biases per dataset. Across five reward models it improves three benchmarks by 8.3, 18.0, and 6.9 percentage points on average, outperforming training-based baselines and reducing downstream verbosity and sycophancy.
This paper analyzes how localizing edits in human corrections can amplify relative selection bias in model training, providing theoretical insights and empirical evidence from machine translation experiments.
DOW-KE introduces an anchor-free method for knowledge editing in large language models by directly optimizing weight updates across layers, achieving higher performance in sequential editing tasks.
Roomer is a reflective repair framework that identifies and fixes local violations in 3D indoor layouts, using a vision-language model planner and deterministic solver, with new benchmarks Roomer-CC and Roomer-Eval.
Introduces V-Steer, a training-free inference-time method that edits cached value vectors to restore instruction hierarchy in language models, raising primary constraint accuracy from under 18% to 92% on controlled benchmarks with negligible overhead.
Anthropic published research on GRAM, a technique for surgically removing dangerous knowledge from AI models at the weight level, advancing AI safety.
This paper introduces process sidecars, a two-coefficient edit family for revoking learned state from language models after safety training, achieving second-order accuracy and outperforming naive task arithmetic in experiments across multiple models.
BrainSurgery is a tool for reproducible and declarative weight manipulations on neural network checkpoints, enabling model editing and upcycling through YAML plans with built-in validation.
NVIDIA introduces Gated DeltaNet-2, a method for editing compressed model memory without catastrophic forgetting, using independent gates for erase and write operations. It outperforms existing models like Mamba-2 and Mamba-3 on language modeling and long-context tasks.
Proposes M-ORE, a modality-decoupled online recursive editor for lifelong adaptation of multimodal large language models, addressing cross-modal conflict and inter-edit interference with constant per-edit overhead.
This paper introduces HoReN, a parameter-preserving model editing method that uses normalized Hopfield retrieval to handle large-scale sequential updates to large language models. It addresses issues of knowledge accumulation and routing challenges, demonstrating stable performance on 50K sequential edits where prior methods degrade.