China's AI-driven Legal-Tech market is rapidly expanding in H1 2026, reaching 4.8 billion RMB with 78% growth, fueled by large AI models and supportive regulatory policies. Key segments include contract review, legal retrieval, and indictment generation, with domestic LLMs achieving high accuracy.
This content introduces "ExpertPrompting," a novel method for instructing Large Language Models to behave as distinguished experts. It focuses on enhancing the specialized knowledge and performance of AI models through advanced prompting techniques.
This content critiques the current assumption that AI memory is a storage problem, proposing an alternative architecture inspired by human memory specialization. It introduces the Personal Small Model (PSM), a small model trained to master memory operations like relevance gating.
Anthropic launched Claude Opus 4.8, introducing dynamic workflows that enable hundreds of parallel subagents for complex tasks. This version shows significant improvements in benchmarks like SWE-bench Verified and USAMO, with unchanged standard pricing and a new, more affordable fast mode.
This content addresses a common mistake in Claude-powered app development: sending the full conversation history with every request, leading to high token costs. It proposes a token-efficient pattern to manage conversation history, ensuring functionality while controlling API spend.
This content explores the fascinating area of translating the internal processes or "thoughts" of an artificial intelligence model like Claude into understandable language. It investigates how the complex operations of AI can be interpreted and expressed to better understand its reasoning.
Together AI is offering free API access to advanced models like Llama 3.3, DeepSeek R1, and FLUX for image generation. This free access will be available until 2026, allowing developers to utilize these powerful AI capabilities without cost.
The author built an AI memory system for Claude that has been running since February. This retrospective explores how the system became self-sufficient and integrated, reducing the need for constant maintenance and intervention from the creator.
Self-Distillation Zero (SD-Zero) is a novel post-training method designed to be more training sample-efficient than traditional reinforcement learning, without requiring external teachers or high-quality demonstrations. It operates by having a single model act as both a Generator and a Reviser, using the Reviser's improved responses and token distributions to provide dense supervision for the Generator through on-policy self-distillation.
This research systematically investigates the Identifiable Victim Effect (IVE) in Large Language Models, a cognitive bias where specific, narratively described victims receive more resources than statistically characterized groups. The large-scale empirical study across 16 frontier LLMs determines if these systems inherit human affective irrationalities in critical applications like humanitarian triage and content moderation.
This paper details an attempt to distill behavioral dispositions into small language models (0.6B-2.3B parameters) through a distillation pipeline. Initial reported gains were later falsified due to evaluation artifacts, resulting in a negative outcome for the core hypothesis and leading to three subsequent arcs of investigation.
This research analyzes Supervised Fine-Tuning (SFT), revealing that instruction-following capabilities emerge distinctly across layers: middle layers are stable while final layers are highly sensitive. Leveraging this, the authors propose Mid-Block Efficient Tuning, which updates critical intermediate layers, outperforming standard LoRA with reduced parameter overhead.
This work introduces an innovative framework for adaptive test-time compute allocation, jointly adjusting where computation is spent and how generation is performed. The method uses a warm-up phase to identify easy queries and then concentrates further computation on unresolved queries, reshaping generation distributions with evolving in-context demonstrations.
This research investigates using an agentic language model feedback framework to generate high-quality planning domains from augmented natural language descriptions. It evaluates the impact of various symbolic feedback mechanisms, like landmarks and plan validation output, in conjunction with heuristic search over model space to optimize domain quality.
To address LLMs' susceptibility to failures from small prompt shifts, especially in multi-step reasoning, researchers propose Distributionally Robust Token Optimization (DRTO). This approach combines token-level Reinforcement Learning from Human Feedback (RLHF) with Distributionally Robust Optimization (DRO) to enhance consistency under distribution shifts, showing improvements on mathematical reasoning benchmarks.
This paper introduces the Length Value Model (LenVM), a novel token-level framework for modeling the remaining generation length in autoregressive models. By formulating length modeling as a value estimation problem, LenVM provides an annotation-free, scalable, and effective signal for LLMs and VLMs, improving performance on exact length matching tasks.
Large language models are often overconfident, expressing high certainty even when incorrect. This paper introduces SECL, a test-time training pipeline that exploits a self-supervised signal to improve calibration without requiring labeled data or human supervision.
This content introduces a novel concept, 'Background Temperature', to characterize the hidden randomness present in Large Language Models.
SpecTr-GBV is a novel speculative decoding method that unifies multi-draft and greedy block verification to accelerate language model inference. It formulates the verification step as an optimal transport problem, improving both theoretical efficiency and empirical performance by achieving the optimal expected acceptance length.
Este artigo enquadra a alucinação em grandes modelos de linguagem como um erro de classificação e propõe uma intervenção composta por recusa baseada em instruções e um gate de abstenção estrutural. O gate utiliza um score de déficit de suporte de sinais como auto-consistência e cobertura de citação, mas a avaliação controlada mostrou que nenhum mecanismo isolado foi suficiente para mitigar totalmente o problema.