20260706:新增一些文章

raw/papers/DSpark-2026.md

@@ -0,0 +1,11 @@
+# Raw Archive: DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation
+
+- **Authors**: Xin Cheng, Xingkai Yu, Chenze Shao, Jiashi Li, Yunfan Xiong, Yi Qian, Jiaqi Zhu, Shirong Ma, Xiaokang Zhang, Jiasheng Ye, Qinyu Chen, Chengqi Deng, Jiping Yu, Damai Dai, Zhengyan Zhang, Yixuan Wei, Yixuan Tan, Wenkai Yang, Runxin Xu, Yu Wu, Zhean Xu, Xuanyu Wang, Muyang Chen, Rui Tian, Xiao Bi, Zhewen Hao, Shaoyuan Chen, Huanqi Cao, Wentao Zhang, Anyi Xu, Huishuai Zhang, Dongyan Zhao, Wenfeng Liang
+- **Affiliations**: Peking University; DeepSeek-AI
+- **Source**: https://github.com/deepseek-ai/DeepSpec/blob/main/DSpark_paper.pdf
+- **Year**: 2026
+- **Code**: https://github.com/deepseek-ai/DeepSpec
+
+## Abstract
+
+Speculative decoding accelerates Large Language Model (LLM) inference by decoupling draft generation from target verification. While recent parallel drafters efficiently propose long token sequences in a single forward pass, they suffer from rapid acceptance decay due to a lack of inter-token dependencies. Furthermore, indiscriminately verifying these extended blocks wastes critical batch capacity on tokens with high rejection risks, severely degrading throughput in high-concurrency serving systems. We introduce DSpark, a speculative decoding framework that unifies high-throughput parallel generation with adaptive, load-aware verification. To maintain draft quality, DSpark utilizes a semi-autoregressive architecture—coupling a parallel backbone with a lightweight sequential module—to introduce intra-block dependency modeling and mitigate suffix decay. To optimize system efficiency, DSpark employs confidence-scheduled verification, dynamically tailoring the verification length for each request based on estimated prefix survival probabilities and engine-specific throughput profiles. On offline benchmarks across diverse domains, DSpark substantially improves the accepted length over state-of-the-art autoregressive and parallel drafters. When deployed within the DeepSeek-V4 serving system under live user traffic, DSpark successfully mitigates verification waste. Compared to the established production baseline (MTP-1), DSpark accelerates per-user generation speeds by 60%–85% at matched throughput levels. More importantly, by preventing severe throughput degradation under strict interactivity constraints, it enables performance tiers that were previously unattainable, shifting the Pareto frontier of our serving system. To facilitate community progress, we open-source the DSpark checkpoints alongside DeepSpec, an algorithm-driven training repository for speculative decoding.

raw/papers/GR4AD-2026.md

@@ -0,0 +1,12 @@
+# Raw Archive: GR4AD — Generative Recommendation for Large-Scale Advertising
+
+- **Authors**: Ben Xue, Dan Liu, Lixiang Wang, Mingjie Sun, Peng Wang, Pengfei Zhang, Shaoyun Shi, Tianyu Xu, Yunhao Sha, Zhiqiang Liu, Bo Kong, Bo Wang, Hang Yang, Jieting Xue, Junhao Wang, Shengyu Wang, Shuping Hui, Wencai Ye, Xiao Lin, Yongzhi Li, Yuhang Chen, Zhihui Yin, Quan Chen, Shiyang Wen, Wenjin Wu, Han Li, Guorui Zhou, Changcheng Li, Peng Jiang (corresponding), Kun Gai
+- **Affiliation**: Kuaishou Technology, Beijing
+- **arXiv**: 2602.22732
+- **Source**: https://arxiv.org/abs/2602.22732
+- **Date**: 2026-02-26 (v3: 2026-04-02)
+- **Status**: Under review
+
+## Abstract
+
+Generative recommendation has recently attracted widespread attention in industry due to its potential for scaling and stronger model capacity. However, deploying real-time generative recommendation in large-scale advertising requires designs beyond large-language-model (LLM)-style training and serving recipes. We present a production-oriented generative recommender co-designed across architecture, learning, and serving, named GR4AD (Generative Recommendation for ADvertising). As for tokenization, GR4AD proposes UA-SID (Unified Advertisement Semantic ID) to capture complicated business information. Furthermore, GR4AD introduces LazyAR, a lazy autoregressive decoder that relaxes layer-wise dependencies for short, multi-candidate generation, preserving effectiveness while reducing inference cost, which facilitates scaling under fixed serving budgets. To align optimization with business value, GR4AD employs VSL (Value-Aware Supervised Learning) and proposes RSPO (Ranking-Guided Softmax Preference Optimization), a ranking-aware, list-wise reinforcement learning algorithm that optimizes value-based rewards under list-level metrics for continual online updates. For online inference, we further propose dynamic beam serving, which adapts beam width across generation levels and online load to control compute. Large-scale online A/B tests show up to 4.2% ad revenue improvement over an existing DLRM-based stack, with consistent gains from both model scaling and inference-time scaling. GR4AD has been fully deployed in Kuaishou advertising system with over 400 million users and achieves high-throughput real-time serving.

raw/papers/leap-agentic-atp.md

@@ -0,0 +1,44 @@
+# LEAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks
+
+- **arXiv ID**: 2606.03303
+- **标题**: LEAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks
+- **作者**: Po-Nien Kung, Linfeng Song, Dawsen Hwang, Jinsung Yoon, Chun-Liang Li, Simone Severini, Mirek Olšák, Edward Lockhart, Quoc V Le, Burak Gokturk, Thang Luong, Tomas Pfister, Nanyun Peng
+- **机构**: Google Cloud AI Research, Google Cloud, Google DeepMind
+- **领域**: cs.AI
+- **发表**: 2026-06-02
+- **代码**: https://github.com/google-deepmind/superhuman/tree/main/leap
+- **Benchmark**: https://imobench.github.io
+
+## 摘要
+
+LEAP (LLM-in-Lean Environment Agentic Prover) 是一个仅使用通用 LLM（无需专用证明器模型微调）的 agentic 形式化定理证明框架。核心设计：(1) 非正式蓝图 + 形式化证明的双层规划；(2) AND-OR DAG 分层记忆化；(3) 验证引导的证明搜索。
+
+## 核心架构
+
+**工作流**：给定定理 → 注册为根目标（OR 节点）→ 直接形式化尝试 → 失败则进入分解路径：
+1. **直接形式化路径**：生成非正式证明 → 翻译为 Lean 代码 → 编译器验证 → 失败则进入 LLM 驱动的修订循环
+2. **分解路径**：生成非正式蓝图（DAG 子目标）→ 翻译为 Lean 证明草图 → 添加为 AND 节点 → 子目标变为新 OR 节点 → 递归处理
+
+**三个关键设计选择**：
+- DAG 分层记忆化：保留进展、跨分支复用引理、支持预期引理规划
+- 非正式-形式化交错规划：LLM 负责推理策略，Lean 负责严格验证
+- 验证引导搜索：编译器形式检查 + LLM reviewer 质量评估（过滤弱分解）
+
+## 关键结果
+
+| Benchmark | LEAP | Aristotle (IMO Gold) | Baseline |
+|-----------|------|---------------------|----------|
+| Putnam 2025 (12 题) | **12/12 (100%)** | 9/12 (75%) | 0/12 (0%) |
+| Lean-IMO-Bench Basic | **70%** | - | <10% |
+| Lean-IMO-Bench Advanced | 显著提升 | 48% (总体) | ~3% |
+
+- 仅 2 次 rollout（vs baselines 的 128 次）
+- 自主形式化 Knuth 哈密顿分解的开放组合挑战的子问题验证证明
+
+## 核心概念
+
+- AND-OR DAG 分层记忆化
+- 非正式蓝图 → 形式化证明交错规划
+- 验证引导证明搜索（编译器 + LLM reviewer）
+- 预期引理规划（anticipatory lemma planning）
+- Lean-IMO-Bench 基准

raw/papers/ramsey-sphere-lowerbound.md

@@ -0,0 +1,65 @@
+---
+title: "An exponential improvement for Ramsey lower bounds (Raw)"
+source: https://arxiv.org/abs/2507.12926
+authors: Jie Ma, Wujie Shen, Shengjie Xie
+institutions: USTC, Tsinghua University, Yau Mathematical Sciences Center
+arxiv: 2507.12926v2
+category: math.CO
+date: April 28, 2026
+---
+
+# An Exponential Improvement for Ramsey Lower Bounds
+
+## Abstract
+
+We prove a new lower bound on the Ramsey number r(ℓ, Cℓ) for any constant C > 1 and sufficiently large ℓ, showing that there exists ε = ε(C) > 0 such that r(ℓ, Cℓ) ≥ (M_C + ε)^ℓ, where M_C = p_C^{-1/2}, and p_C ∈ (0, 1/2) is the unique solution to C = log p_C / log(1-p_C). This provides the first exponential improvement over the classical lower bound obtained by Erdős in 1947.
+
+## 1. Introduction
+
+The Ramsey number r(ℓ, k) denotes the smallest positive integer n such that every red-blue edge coloring of the complete graph K_n on n vertices contains either a red clique K_ℓ or a blue clique K_k.
+
+**History of upper bounds**:
+- 1935: Erdős-Szekeres — r(ℓ, k) ≤ C(k+ℓ-2, ℓ-1)
+- 1988: Thomason — first polynomial improvement
+- 2009: Conlon — superpolynomial improvement via quasi-randomness
+- 2013: Sah — refined optimization
+- 2023: Campos, Griffiths, Morris, Sahasrabudhe — first exponential improvement: r(ℓ, k) ≤ e^{-ℓ/400+o(k)}·C(k+ℓ, ℓ)
+- 2025: Gupta, Ndiaye, Norin, Wei — improved to r(ℓ, ℓ) ≤ 3.8^{ℓ+o(ℓ)}
+
+**History of lower bounds**:
+- 1947: Erdős — probabilistic method: r(ℓ, Cℓ) = Ω(ℓ·M_C^ℓ)
+- 1975: Spencer — Lovász Local Lemma constant-factor refinement
+- 2026: **This paper** — first exponential improvement
+
+## Theorem 1.1 (Main Result)
+
+For any constant C > 1, there exist ε = ε(C) > 0 and ℓ₀ = ℓ₀(C) > 0 such that for all ℓ ≥ ℓ₀(C),
+
+r(ℓ, Cℓ) ≥ (M_C + ε)^ℓ
+
+where M_C = p_C^{-1/2}, and p_C ∈ (0, 1/2) satisfies C = log p_C / log(1-p_C).
+
+## Corollary 1.2 (General Regime)
+
+For any δ ∈ (0, 1/2), r(ℓ, k) ≥ (1+2c_δ)^ℓ · (M_{k/ℓ})^ℓ ≥ (1+c_δ)^ℓ · Er(ℓ, k) whenever δ ≤ ℓ/k ≤ 1-δ.
+
+## Corollary (Almost Diagonal)
+
+For √ℓ ≪ f(ℓ) ≪ ℓ: r(ℓ, ℓ+f(ℓ)) ≥ e^{Ω(f(ℓ)²/ℓ)} · Er(ℓ, ℓ+f(ℓ))
+
+## 2. The Random Sphere Graph G_{k,p}(n)
+
+A novel random graph model based on geometric measure:
+- Sample n points uniformly at random from the k-dimensional unit sphere S^k ⊂ R^{k+1}
+- Connect each pair independently with probability p
+- Unlike G(n,p), the edge probability is governed by geometric proximity
+
+## 3-9. Proof Structure
+
+1. **Section 3**: Reduce Theorem 1.1 to Theorem 3.1 (core technical result)
+2. **Section 4**: Auxiliary lemmas (geometric measure estimates)
+3. **Section 5**: Introduce perfect sequences for unit vectors
+4. **Section 6**: Preliminary estimates on perfect sequences
+5. **Section 7**: Show perfect sequences capture essential behavior
+6. **Section 8**: Core technical arguments — estimates on key quantities
+7. **Section 9**: Assemble all estimates to complete proof of Theorem 3.1

raw/papers/rubrics-survey-2026.md

@@ -0,0 +1,54 @@
+---
+title: "The Rules of the Game: A Survey of Rubrics for Large Language Models"
+created: 2026-06-27
+updated: 2026-06-27
+type: raw-paper
+source_url: "https://8421bcd.github.io/_pages/Rubrics_Survey.pdf"
+github: "https://github.com/8421BCD/Rubrics_Survey"
+authors:
+  - "Wenhan Liu"
+  - "Jiajie Jin"
+  - "Zhaoheng Huang"
+  - "Tongyu Wen"
+  - "Guanting Dong"
+  - "Ziliang Zhao"
+  - "Yutao Zhu"
+  - "Zhicheng Dou"
+  - "Ji-Rong Wen"
+affiliation: "Renmin University of China"
+date: "2026-05-22"
+tags:
+  - rubric
+  - evaluation
+  - reward-modeling
+  - survey
+---
+
+# The Rules of the Game: A Survey of Rubrics for Large Language Models
+
+## Abstract
+
+As LLMs evolve from general text generators into systems for reasoning, decision-making, tool use, and long-horizon problem solving, the question of how to specify, optimize, and evaluate model responses has become critical. Simple correctness signals and unconstrained LLM-based judgments are often insufficient for open-ended and high-stakes tasks where response quality depends on multiple criteria (factuality, completeness, safety, reasoning soundness, evidence grounding, practical utility).
+
+Rubrics have emerged as a mechanism for making evaluation standards explicit and operational. This survey:
+
+1. **Defines rubrics** and distinguishes them from reward models, verifiable rewards, and LLM-as-a-judge
+2. **Organizes rubric construction** into four paradigms: direct generation, contrastive generation, iterative refinement, and online/co-evolving generation
+3. **Examines rubric-driven training** for both policy models (standard RL, advanced reward design, policy guidance) and reward models (interpretability, reward signals, data construction)
+4. **Summarizes rubric-driven evaluation** across general tasks (reasoning, deep research, agent capability, alignment) and domain-specific tasks (intermediate trajectories, final outputs)
+5. **Discusses open challenges**: rubric reward hacking, generalization of rubric-based reward models, bias in rubric-based evaluation, personalized rubrics, and rubric safety
+
+## Key Contributions
+
+- First comprehensive survey on rubrics for LLMs
+- Systematic taxonomy of rubric construction methods (4 categories)
+- Comprehensive review of rubrics for model training and evaluation
+- Discussion of open challenges and future directions (personalized rubrics, rubric safety, RIPD)
+
+## Formal Definition
+
+A rubric is a set of rubric items: R = {(dⱼ, wⱼ)}ᵏⱼ₌₁
+- dⱼ: natural-language description of the j-th rubric item
+- wⱼ: its weight
+- Score: cⱼ(x, y) ∈ [0, 1]
+- Aggregated rubric score: S_R(x, y) = Σwⱼcⱼ(x,y) / Σwⱼ

raw/papers/safe-equilibrium-exploration.md

@@ -0,0 +1,61 @@
+---
+title: "Safe Equilibrium Exploration (Raw)"
+source: https://arxiv.org/abs/2602.00636
+authors: Yujie Yang, Zhilong Zheng, Shengbo Eben Li
+institution: Tsinghua University, School of Vehicle and Mobility
+arxiv: 2602.00636v2
+category: cs.LG
+venue: IEEE TPAMI 48(7), 8344-8360 (2026)
+doi: 10.1109/TPAMI.2026.3669907
+date: January 31, 2026
+---
+
+# On the Equilibrium between Feasible Zone and Uncertain Model in Safe Exploration
+
+## Abstract
+
+Ensuring the safety of environmental exploration is a critical problem in RL. While limiting exploration to a feasible zone has become widely accepted, key questions remain: what is the maximum feasible zone achievable through exploration, and how can it be identified? This paper reveals that the goal of safe exploration is to find the equilibrium between the feasible zone and the environment model. These two components are interdependent: a larger feasible zone leads to a more accurate environment model, and a more accurate model enables exploring a larger zone. We propose Safe Equilibrium Exploration (SEE), which alternates between finding the maximum feasible zone and the least uncertain model. Using a graph formulation, we prove the uncertain model is monotonically refined, feasible zones monotonically expand, and both converge to the equilibrium. Experiments on classic control tasks show zero constraint violation and equilibrium within a few iterations.
+
+## 1. Introduction
+
+**Training modes for safe RL**:
+- **OTOI** (Offline Training Online Implementation): Train in simulator, deploy online — no safe exploration needed, but requires high-fidelity sim
+- **SOTI** (Simultaneous Online Training Implementation): Learn through online interaction — requires safe exploration
+
+**Key insight**: Safe exploration isn't about maximizing feasible zone alone, but finding equilibrium with the uncertain model.
+
+## 2. Related Work
+
+**Safety filter methods**:
+- Dalal et al. (2018): Action correction under instantaneous constraints
+- Pham et al. (2018): QP layer for distance/velocity constraints
+- Cheng et al. (2019): State-affine CBF + QP safety filter
+- Control Barrier Function (CBF) based methods
+- Safety Index method (Liu et al.)
+
+**Limitation**: Safety filter methods rely on human-designed constraints → feasible zones are conservative and incomplete.
+
+## 3. Problem Formulation
+
+- MDP with safety constraints
+- Feasible zone Z: subset of state-action space where constraint-satisfying policy exists
+- Uncertain model M: estimated transition dynamics with quantified uncertainty
+- Equilibrium condition: Z is the maximum feasible zone given M, and M is the least uncertain model given data collected within Z
+
+## 4. SEE Algorithm
+
+Alternating optimization:
+1. **Zone expansion**: Given current model M, find maximum feasible zone Z
+2. **Model refinement**: Collect data within Z, update model M
+3. Repeat until convergence
+
+## 5. Theoretical Analysis
+
+- Monotonic refinement of uncertain model
+- Monotonic expansion of feasible zones
+- Convergence to equilibrium
+- Graph formulation as the analytical framework
+
+## 6. Experiments
+
+Classic control tasks: SEE achieves zero constraint violation with monotonic zone expansion, reaching equilibrium within few iterations.

raw/papers/sen-mapping-networks-2026.md

@@ -0,0 +1,49 @@
+# Mapping Networks — Raw Archive
+
+- **Title**: Mapping Networks
+- **Authors**: Lord Sen, Shyamapada Mukherjee (National Institute of Technology Rourkela, India)
+- **arXiv ID**: 2602.19134
+- **Category**: cs.CV (Computer Vision)
+- **Date**: 2026-02-22
+- **Pages**: 10
+
+## Abstract
+
+The escalating parameter counts in modern deep learning models pose a fundamental challenge to efficient training and resolution of overfitting. We address this by introducing the *Mapping Networks* which replace the high dimensional weight space by a compact, trainable latent vector based on the hypothesis that the trained parameters of large networks reside on smooth, low-dimensional manifolds. Henceforth, the Mapping Theorem enforced by a dedicated Mapping Loss, shows the existence of a mapping from this latent space to the target weight space both theoretically and in practice. Mapping Networks significantly reduce overfitting and achieve comparable to better performance than target network across complex vision and sequence tasks, including Image Classification, Deepfake Detection etc, with 99.5%, i.e., around 500× reduction in trainable parameters.
+
+## Key Contributions
+
+1. **Mapping Theorem**: Establishes existence of a smooth, low-dimensional parameterization capable of generating optimal high-dimensional weights for a target network with arbitrarily small bounded error.
+2. **Mapping Network**: Low-dimensional trainable latent vector + modulated Mapping Weights to efficiently produce target network parameters, decoupling training from target network.
+3. **Mapping Loss**: Jointly optimizes task performance and enforces geometric/analytic properties required by Mapping Theorem (Task + Stability + Smoothness + Alignment).
+
+## Core Mechanism
+
+- **Weight-Manifold Hypothesis**: Neural network parameters during training do not explore full P-dimensional Euclidean space; they evolve along smooth, low-dimensional manifolds M_θ ⊂ R^P with dim(M_θ) ≪ P.
+- **Architecture**: Trainable latent vector z ∈ R^d → fixed orthogonally initialized mapping network modulated by z (w_ij ← w_ij + α·z_i) → generates target network parameters θ̂ → target network does feed-forward only.
+- **Mapping Loss**: L_map = L_task + λ·L_stab + λ·L_smooth + λ·L_align (all λ trainable)
+- **Two strategies**: SLVT (single latent vector) and LWT (layer-wise training)
+
+## Results Summary
+
+| Task | Baseline Params | Mapping Params | Reduction | Performance |
+|------|----------------|----------------|-----------|-------------|
+| MNIST (CNN1) | 537,994 | 1,024-4,078 | 260-525× | 98.78-99.67% (vs 99.32%) |
+| FMNIST (CNN1) | 537,994 | 1,024-4,078 | 260-525× | 93.02-94.83% (vs 92.89%) |
+| Deepfake (CNN2, Celeb-DF) | 108,618 | 1,024-2,688 | 40-106× | 78.83-89.98% (vs 79.03-83.13%) |
+| Segmentation (CNN3) | 1,734,803 | 8,192-9,126 | 190-211× | 97.56-97.92% (vs 93.21%) |
+| LSTM (Air Pollution) | 12,961 | 64-2,048 | 6-202× | MSE 0.00061-0.0019 (vs 0.0035) |
+| Fine-tune ResNet50 | 25M | 1,024-2,048 | ~12,000× | 91.02-95.10% (vs 91.78-95.23%) |
+
+## Related Work
+
+- Manifold Hypothesis [7, 19]
+- HyperNetworks [13]
+- Low-Rank Compression [14]
+- Lottery Ticket Hypothesis [9]
+- Intrinsic Dimension of loss landscapes [16]
+- Training trajectory manifolds [18]
+
+## Extensions
+
+Compatible with Low-Rank Decomposition, Pruning, and Quantization for further inference-time optimization.

raw/papers/tapered-language-models.md

@@ -0,0 +1,60 @@
+---
+title: "Tapered Language Models (Raw)"
+source: https://arxiv.org/abs/2606.23670
+authors: Reza Bayat, Ali Behrouz, Aaron Courville
+institutions: Mila, Cornell University, Université de Montréal, CIFAR
+arxiv: 2606.23670v1
+category: cs.LG
+date: June 22, 2026
+---
+
+# Tapered Language Models
+
+## Abstract
+
+Modern LMs share a common chassis: a stack of identical layers with parameters allocated uniformly across depth. Evidence suggests layers contribute non-uniformly—later layers refine rather than transform the residual stream. We ask: should parameter capacity reflect this asymmetry? Under fixed budget, allocating more capacity to earlier layers and less to later layers improves perplexity; reverse allocation hurts. We introduce Tapered Language Models (TLMs), where a parameter-bearing component is monotonically tapered across depth. MLPs are the natural site: they dominate parameter count and expose width as a clean axis. Across three scales and four architectures, tapering MLP width via cosine schedule consistently improves perplexity and downstream benchmarks at no additional cost.
+
+## 1. Introduction
+
+Uniform layer width is an inherited default from Vaswani et al. (2017). As models scaled, this uniformity remained unexamined.
+
+## 2. Related Work
+
+- Mixture-of-Experts (MoE): conditional parameter allocation
+- LayerDrop, stochastic depth: layer-wise computation pruning
+- Depth-wise scaling: changing number of layers
+
+## 3. Tapered Language Models
+
+**Principle**: Under fixed total parameter budget, monotonically decrease parameter allocation from early to late layers.
+
+**Why MLP width**: 
+- MLP accounts for majority of parameters in Transformer, Gated Attention, Mamba, Titans
+- Width (d_ff) is a single clean axis of variation
+- Token-mixing modules vary across architectures, making uniform comparison difficult
+
+**Cosine schedule**: w_ℓ = w_max · (cos(πℓ/(2L)))^p, where p controls steepness
+
+## 4. Experiments
+
+**Controlled study** (440M Transformer):
+- Uniform: 16.28 perplexity
+- Cosine taper (1.50→0.50 × d_ff): 14.44 perplexity
+- Cosine dominates linear at every taper range
+- U-shape: optimal at 1.50→0.50, extreme tapers underperform
+
+**Architecture sweep** (440M):
+- Transformer ✓
+- Gated Attention ✓ 
+- Hope-attention ✓
+- Titans ✓
+
+**Scale sweep**: 440M, 1B, 3B — consistent improvement
+
+**Downstream**: HellaSwag, ARC-E, PIQA, WinoGrande — taper improves all
+
+## 5. Analysis
+
+- Early layers benefit more from MLP capacity (transformation)
+- Late layers benefit less (refinement of residual stream)
+- Cosine > Linear > Step-wise > Uniform

raw/papers/verification-horizon-qwen-2026.md

@@ -0,0 +1,35 @@
+---
+title: "The Verification Horizon: No Silver Bullet for Coding Agent Rewards"
+created: 2026-07-02
+type: raw
+arxiv: "2606.26300"
+authors: "Binghai Wang, Chenlong Zhang, Dayiheng Liu, Jiajun Zhang, Jiawei Chen, Mingze Li, Mouxiang Chen, Rongyao Fang, Siyuan Zhang, Xuwu Wang, Yuheng Jing, Zeyao Ma, Zeyu Cui (Qwen Team, Alibaba)"
+venue: arXiv
+date: 2026-06-24
+---
+
+# Raw: The Verification Horizon
+
+A classical intuition holds that verifying a solution is easier than producing one. For today's coding agents, this intuition is being inverted: generating complex candidate solutions is no longer difficult — reliably verifying them has become the harder problem.
+
+## Core Framework
+
+Every verifier is a proxy for human intent, never the intent itself. Verification quality characterized along three dimensions:
+- **Scalability**: can the signal be produced cheaply at training scale?
+- **Faithfulness**: how much of true user intent does the signal reflect?
+- **Robustness**: can the verifier hold across diverse inputs and optimization pressure?
+
+Achieving all three simultaneously is the central challenge.
+
+## Four Reward Constructions
+
+1. **Test Verifier** (§2): execution-based test suites + agentic quality judge + behavior monitoring. Hacked resolved rate: 28.57% → 0.56%; Clean resolved: 40.22% → 60.53%.
+2. **Interactive Judge** (§3): rubric-based static judge → agentic interactive judge with Playwright. Resists length exploitation.
+3. **User Feedback Verifier** (§4): extract HIRS from interaction data → Span-KTO training. +13.3pp on Aone-bench.
+4. **Automated Agent Verifier** (§5): autonomous evaluator for long-horizon tasks. Evaluator quality is metric-dependent.
+
+## Core Claim
+
+No fixed reward function can remain effective as policy capability continues to grow; verification must co-evolve with the generator.
+
+Source: https://arxiv.org/abs/2606.26300

raw/papers/yang-semantic-robustness-cert-2026.md

@@ -0,0 +1,25 @@
+---
+title: "Semantic Robustness Certification for Vision-Language Models (Raw)"
+created: 2026-07-04
+type: raw-paper
+source: "arXiv:2606.18839"
+venue: "ICML 2026"
+authors: ["Peiyu Yang", "Paul Montague", "Feng Liu", "Andrew C. Cullen", "Amardeep Kaur", "Christopher Leckie", "Sarah M. Erfani"]
+---
+
+# Semantic Robustness Certification for Vision-Language Models
+
+- **arXiv**: 2606.18839
+- **Venue**: ICML 2026
+- **Institutions**: University of Melbourne, Defence Science and Technology Group
+- **Code**: https://github.com/ypeiyu/vlm-semantic-cert
+
+## Abstract
+
+Vision-language models (VLMs) are now widely used in downstream tasks. However, real-world applications often expose VLMs to distribution shifts induced by semantic variation (e.g., shape, size, and style). Robustness certification determines if a model's prediction changes when transformations are applied to its input. While most certification frameworks study geometric or pixel-level transformations over inputs, this work proposes a novel framework that enables certifying VLM robustness under semantic-level transformations. Leveraging the open-vocabulary capability of VLMs, we use text prompts as semantic proxies to construct transformations parameterized by an extent that controls the degree of semantic variation. By characterizing the VLM decision boundary in closed form, our framework quantitatively certifies extent intervals for which the predicted class remains unchanged under the semantic transformation. Our framework is the first to certify VLM robustness under semantic-level variations without requiring additional data for each variation, making it practical to apply.
+
+## Key Contributions
+
+1. Text prompts as semantic proxies to formalize semantic transformations for VLMs
+2. Closed-form characterization of VLM decision boundary → precise prediction-invariant intervals
+3. Evaluations on both synthetic and real-world data — transformations align with target semantics, certificates match prediction changes