20260720

raw/articles/openai-gpt5p6-prompt-guide-2026.md

@@ -0,0 +1,35 @@
+# OpenAI GPT-5.6 官方 Prompt 指南解读
+
+- **来源**: Datawhale 微信公众号
+- **链接**: https://mp.weixin.qq.com/s/lSvGH3nCK9oWf8wOyeCTGA
+- **作者**: Eric Provencher (OpenAI Codex DX 团队)
+- **官方原文**: https://developers.openai.com/api/docs/guides/prompt-guidance-gpt-5p6
+- **日期**: 2026-07
+
+## 核心主题
+
+OpenAI 发布 GPT-5.6 官方 Prompt 指南，核心理念：**模型越来越强，Prompt 应该越来越精简**。旧模型留下的补丁、重复规则和层层叠加的工具说明不仅消耗 Token，还可能互相冲突。
+
+## 关键数据
+
+在 OpenAI 内部 Coding Agent 任务中，精简后的 System Prompt：
+- 评分提高 10%-15%
+- Token 总量减少 41%-66%
+- 成本降低 33%-67%
+
+## 核心原则
+
+1. **Prompt 做减法**：删无效信息，留必要要求。每次只删一组，用同一批评测验证
+2. **结果优先**：写清最终目标、可用证据、行动边界和验收标准，而非规定每一步
+3. **明确授权边界**：分析/审查 ≠ 修改/构建，外部写入/删除需再次确认
+4. **停止条件前置**：证据足够→交付，缺少事实→低成本补充
+5. **交付前验证**：代码→测试+lint+构建检查；前端→实际渲染检查
+6. **推理强度调优**：先检查 Prompt 质量再提升 Reasoning Effort
+
+## Prompt 结构框架
+
+Role → Personality → Goal → Success criteria → Constraints → Tools → Output → Stop rules
+
+## 关键词
+
+prompt engineering, GPT-5.6, system prompt, token optimization, agent development

raw/papers/Cheng-etal-SMG-2026.md

@@ -0,0 +1,20 @@
+# Statistically Meaningful Geometry (SMG) Beyond the Euclidean Paradigm (Raw Archive)
+
+- **arXiv**: 2607.03329v1
+- **Authors**: Bing Cheng (CAS), Yi-Shuai Niu (BIMSA), Howell Tong (LSE/Tsinghua), **Shing-Tung Yau** (BIMSA/Tsinghua)
+- **Published**: July 3, 2026
+- **Category**: cs.LG, stat.ME
+- **Length**: 96 pages
+
+## Abstract
+
+Conventional uniform convergence bounds and empirical risk minimization break down in massive over-parameterized models. With near-infinite unconstrained internal degrees of freedom, optimization landscapes develop flat vertical gauge valleys, rendering classical generalization metrics vacuous and inducing generative hallucination and catastrophic forgetting. We introduce the Statistically Meaningful Geometry (SMG) framework, an information-geometric paradigm lifting deterministic parametric models into infinite-dimensional non-parametric Orlicz statistical manifolds. Modeling the total state space as a differential fiber bundle (M, B, π, V, H, ω), we establish a Two-Fold Inference Paradigm. We formalize an Ehresmann connection 1-form ω as a dynamic geometric filter that strips away vertical gauge noise (SID) and isolates learning along the non-degenerate horizontal distribution (SVDχ). We prove that under connection-filtered pre-training, out-of-distribution predictive variance is strictly upper-bounded by the finite diameter of the identifiable quotient base manifold B — a hard geometric containment of generative hallucinations. By projecting downstream updates onto the orthogonal complement of the historical horizontal carriage, we formalize the SMG Sequential Adaptation Flow, proving total non-asymptotic elimination of catastrophic forgetting.
+
+## Key Framework Elements
+
+- **SMG Fiber Bundle**: (M, B, π, V, H, ω) — total Orlicz manifold M, base manifold B, projection π, vertical subspace V (SID), horizontal distribution H (SVDχ), Ehresmann connection ω
+- **Four Core Axioms**: System Set, Environment Set, Structural Mechanism, Invariance Principle
+- **Two-Fold Inference Paradigm**: Horizontal (statistical on B, classical MLE/Bayes) + Vertical (geometric, connection-driven navigation of SID)
+- **Quarantining Theorem**: Orthogonal metric decomposition Tf M = Hf ⊕ Vf with Fisher-Rao metric orthogonality — vertical gauge noise cannot contaminate statistical signal
+- **PAC-Bayesian Capacity Collapse Theorem**: Generalization bounded by finite metric volume of base space B despite infinite parameters
+- **SMG Sequential Adaptation Flow**: Non-asymptotic catastrophic forgetting elimination via orthogonal projection onto complement of historical horizontal carriage

raw/papers/Hu-etal-concepts-shared-geometry-2026.md

@@ -0,0 +1,19 @@
+# Language Models Represent and Transform Concepts with Shared Geometry (Raw Archive)
+
+- **arXiv**: 2607.04525v1
+- **Authors**: Zhimin Hu (Georgia Tech), Lanhao Niu (U. Edinburgh), Sashank Varma (Georgia Tech)
+- **Published**: July 5, 2026
+- **Venue**: ICML 2026 (Proceedings of the 43rd ICML, Seoul)
+- **Category**: cs.CL, cs.AI
+
+## Abstract
+
+How concepts are represented in neural networks is a fundamental question. The dominant view treats concept representations as stationary geometric objects. Yet concepts appear in context, and context transforms them. Drawing from neural population geometry, we formalize concept representations as point-cloud manifolds and contextual transformations as vector fields, and instantiate this framework in large language models. Across six model families of varying scales, we find that context moves each concept differently. The variance in these displacements is semantically organized, correlating with lexical concreteness and density. Importantly, both the concepts being transformed and this variance structure are shared across models: displacement structure transported from one model predicts held-out displacements in others significantly above chance.
+
+## Key Framework
+
+- **Concepts as Point-Cloud Manifolds**: X_w(m) = {r(m)(w, c_i)} — each concept is an empirical sample from an underlying distribution
+- **Two Axes of Representation**: Within-concept (shape of point cloud across contexts) and Between-concept (kernel matrix of concepts in shared context)
+- **Contextual Transformation Field**: φ(m)(w, τ) = r(m)(w, τ) - r(m)(w, τ_0), characterized by its kernel matrix
+- **Cross-Model Measurement**: CKA for relational structure; Grassmann distance for directional alignment; permutation calibration for significance
+- **23 models across 6 families** (0.5B–32B): Qwen2.5, Llama, Mistral, Gemma2, DeepSeek, Pythia

raw/papers/Levada-CuBAS-curvature-adaptive-sampling-2026.md

@@ -0,0 +1,20 @@
+# CuBAS: Curvature-Based Adaptive Sampling for Supervised Classification (Raw Archive)
+
+- **arXiv**: 2607.03145v1
+- **Author**: Alexandre L. M. Levada (Federal University of São Carlos)
+- **Published**: July 3, 2026
+- **Category**: cs.LG, cs.AI, cs.CV, cs.IT, stat.ML
+- **Code**: https://github.com/alexandrelevada/CuBAS
+
+## Abstract
+
+The informativeness of a training set is as consequential as its size, yet most sampling strategies remain agnostic to the intrinsic geometry of the data distribution. We introduce CuBAS (Curvature-Based Adaptive Sampling), an information-geometric framework for adaptive data selection in supervised classification, grounded in the q-state Potts Markov random field (MRF) model. The central insight is that a labeled dataset can be viewed as a statistical manifold, on which local curvature, estimated via the ratio of second- to first-order observed Fisher information, faithfully encodes the geometric complexity of the underlying data distribution. We construct a k-nearest-neighbor graph over the labeled data and derive a closed-form curvature score at each vertex from the Potts sufficient statistics. This curvature signal partitions the graph into two complementary regimes: low-curvature regions (smooth, homogeneous clusters) and high-curvature regions (decision boundaries, disproportionately informative). By selecting nodes from both regimes, CuBAS constructs compact yet maximally informative training subsets. Extensive evaluation across 60+ benchmark datasets demonstrates consistent and statistically significant improvements over random sampling and uncertainty-based baselines.
+
+## Key Technical Details
+
+- **Potts MRF**: Isotropic q-state Potts model on k-NN graph; local conditional distribution parameterized by inverse temperature β
+- **MPL Estimation**: Maximum pseudo-likelihood with secant method for β
+- **Curvature Score**: S_i(β) = -Ψ_i(β) / (Φ_i(β) + λ), where Φ_i is 1st-order and Ψ_i is 2nd-order observed Fisher information
+- **Adaptive Threshold**: Otsu's criterion → Ashman's D bimodality check → Tukey-fence fallback
+- **Complexity**: O(nm log n + nk + T_β nc), linear in k-NN graph edges after construction
+- **Results**: Mean +13.5pp over entropy sampling at 10% budget across 42 datasets; best on all 60+ datasets

raw/papers/You-GID-sphere-2026.md

@@ -0,0 +1,20 @@
+# Geometric Information Decomposition for Weighted Empirical Measures on the Sphere (Raw Archive)
+
+- **arXiv**: 2607.02943v1
+- **Author**: Kisung You (Baruch College)
+- **Published**: July 3, 2026
+- **Category**: stat.ME
+- **Code**: https://github.com/kisungyou/GeoInfoDec
+
+## Abstract
+
+We study directional uncertainty when the data already represent a weighted probability measure on the unit sphere, as in importance samples, quadrature rules, or attention-weighted embeddings. Standard approach fits a von Mises-Fisher distribution and reports concentration or entropy — principled but incomplete, because vMF uses only mean-direction information and can miss antipodal, axial, girdle-like, or multimodal structure. We introduce the geometric information decomposition (GID), which fits a nested sequence of maximum-entropy projections using spherical features and reports the entropy gap added at each level. First gap recovers vMF information, second captures Fisher-Bingham/Bingham-type anisotropy, and later gaps capture finer angular structure. We prove invariance, consistency, asymptotic normality away from zero gaps, and a quadratic-form null calibration for deciding whether a new level carries information.
+
+## Key Technical Details
+
+- **Feature spaces**: Nested V₀ ⊂ V₁ ⊂ ... ⊂ V_L ⊂ L²₀(M, ν), with V₀ = {0} (uniform)
+- **Maximum-entropy projection**: p_L,m = exp{λ_L(m)ᵀ φ_L(x) - ψ_L(λ_L(m))}, where λ solves ∇ψ_L(λ) = m
+- **Entropy deficit**: D_L(P) = KL(p^P_L ν ∥ ν) — non-uniform info explained up to level L
+- **Information gap**: I_L(P) = D_L(P) - D_{L-1}(P) — additional info at level L
+- **Level 1** = vMF mean-direction, **Level 2** = Fisher-Bingham axial/girdle, **Level 3+** = multimodal
+- **Null calibration**: 2a²_n Î_L ⇒ χ²_q under null I_L=0 (Theorem 8)

raw/papers/agent-harness-survey-2026.md

@@ -0,0 +1,20 @@
+# From Question Answering to Task Completion: A Survey on Agent System and Harness Design
+
+- **arXiv**: 2606.20683v1
+- **Published**: 2026-06-14
+- **Categories**: cs.AI, cs.CL
+- **Authors**: Jianyuan Guo, Zhiwei Hao, Chengcheng Wang, Cheng Fan, Tingzhang Luo, Hongguang Li, Ying Gao, Hefei Mei, Jiankun Peng, Rongjian Xu, Minjing Dong, Han Wu, Mengyu Zheng, Kai Han, Shiqi Wang, Chang Xu, Yunhe Wang
+- **Affiliations**: City University of Hong Kong, University of Sydney, Peking University, TokenRhythm Technologies
+- **Repository**: https://github.com/ggjy/Awesome-Agent-Engineering
+
+## Abstract
+
+LLM-based agents mark a shift from passive question answering to active task completion: they perceive environments, invoke tools, maintain state, and act over extended horizons. As agent systems have evolved from prompt engineering to workflows and context engineering, harness engineering, and agent-native training with co-evolution, a central question has become increasingly important: where does the bottleneck in agent performance reside—in the foundation model, in the execution harness, or in the coupling between them? This survey examines LLM-based agents through a model–harness lens. We first clarify the functional definition of agents and the implementation view of an LLM-based agent as a foundation model coupled with an execution harness. We then analyze the limits of model-centric scaling, trace four paradigms of agent engineering, and decompose the execution harness into six coupled runtime responsibilities: observation, context, control, action, state, and verification/governance. Using this decomposition, we map task properties and domain pressures to harness configurations, review benchmark and evaluation practices, and synthesize model–harness evidence on how runtime design affects long-horizon task completion, efficiency, and reliability. Finally, we identify open challenges in value-aware evaluation, safety, harness generalization, and model–harness co-evolution.
+
+## Key Contributions
+
+1. **Model-Harness Lens**: formal definition Agent = ⟨M, H⟩, where M is model layer and H is execution harness with six components
+2. **Four Paradigms of Agent Engineering**: prompt engineering → context engineering → harness engineering → agent-native training & co-evolution
+3. **Harness Anatomy**: six coupled runtime components — I_obs (observation), C (context), L (control loop), I_act (action), S (state/artifact store), V (verification/governance)
+4. **Harness-Aware Task Taxonomy**: task complexity L1-L4 mapped to harness bottlenecks
+5. **Cross-Layer Interactions**: harness as coupled systems problem, not independent module optimization

raw/papers/genception-video-vision-2026.md

@@ -0,0 +1,21 @@
+# Video Generation Models are General-Purpose Vision Learners (GenCeption)
+
+- **arXiv**: 2607.09024v1
+- **Published**: 2026-07-10
+- **Venue**: ECCV 2026
+- **Categories**: cs.CV, cs.AI
+- **Authors**: Letian Wang, Chuhan Zhang, Rishabh Kabra, Jasper Uijlings, Steven Waslander, Andrew Zisserman, Joao Carreira, Kaiming He, Misha Andriluka, Eduard Gabriel Bazavan, Andrei Zanfir, Cristian Sminchisescu
+- **Affiliations**: Google DeepMind, University of Toronto, UCL, University of Oxford, MIT, Lund University
+- **Project**: https://genception.github.io
+
+## Abstract
+
+The paper argues that large-scale text-to-video generation serves as a strong pre-training paradigm for computer vision — the visual equivalent of next-token prediction in NLP. GenCeption repurposes a pre-trained video generative diffusion backbone into a feed-forward perception model capable of diverse vision tasks steered by text instructions. It achieves SOTA across depth, surface normal, camera pose estimation, expression-referring segmentation, and 3D keypoint prediction, often matching or surpassing specialized models with 7x-500x less training data. Notably, a model trained exclusively on synthetic human videos generalizes to real-world footage and OOD categories (animals, robots).
+
+## Key Contributions
+
+1. **Video Generation as Universal Vision Pre-training**: satisfies three imperatives—spatio-temporal priors, vision-language alignment, and scalability
+2. **Feed-Forward Repurposing**: iterative diffusion → single-step perception via conditioning on t=0 and negating velocity output
+3. **Unified Task Representation**: all dense tasks in 3-channel RGB space; sparse tasks via learnable tokens
+4. **Data Efficiency**: matches SOTA specialists with 7x-500x less training data
+5. **Emergent Sim-to-Real Transfer**: synthetic-only training generalizes to real-world and OOD categories