{
  "docId": "019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "docSlug": "5df6bd1b0447b5f6",
  "documentTitle": "2025 Air Street Capital The State of AI Report 2025",
  "authorId": "AirStreetCapital",
  "authorName": "Air Street Capital",
  "documentKindSlug": "consulting-deck",
  "documentKindLabel": "Consulting deck",
  "sourceTypeSlug": "vc_research",
  "sourceTypeLabel": "VC research",
  "presentationDate": null,
  "orientation": "landscape",
  "aspectRatio": 1.777,
  "pageNumber": 33,
  "pageCount": 313,
  "prevPage": 32,
  "nextPage": 34,
  "slideType": "industry_trends",
  "function": "analyze_data",
  "density": "dense",
  "nDataPoints": 4,
  "notes": "The slide discusses the debate on whether RLVR creates new reasoning or just reshuffles sampling, citing specific research from Tsinghua and MSR Asia.",
  "elementsJson": [
    "headline_text",
    "paragraph",
    "bullet_list",
    "bar_chart_grouped",
    "process_diagram",
    "line_chart"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/33",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-33",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Together they map where RLVR helps today and where it stalls.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a81d-1263e0cec569",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.25,
        "x": 0.7,
        "y": 0.42
      },
      "kind": "chart",
      "text": "Frequency vs Accuracy Interval histogram",
      "attrs": null,
      "subkind": "bar-grouped",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "1e6474f0-4f46-471e-be7c-baa41b6fd4bb",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.3,
        "x": 0.68,
        "y": 0.7
      },
      "kind": "diagram",
      "text": "Conceptual diagram of RLVR reasoning paths and performance charts",
      "attrs": null,
      "subkind": "process",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "92385bae-046d-4478-99b1-416d85563425",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.4,
        "w": 0.65,
        "x": 0.02,
        "y": 0.45
      },
      "kind": "list",
      "text": "Work from Tsinghua evaluated many models, tasks, and RL algorithms and find that present-day RLVR improves Pass@1 but, at larger K, base models catch up. They conclude RLVR has not unlocked fundamentally new reasoning and remains bounded by the base model's capacity.\nA counter from MSR Asia formalized why Pass@K can mask progress and introduce CoT-Pass@K, which requires both a correct answer and a valid chain-of-thought.\nOn AIME-2024/2025 with Qwen2.5-32B -> DAPO-Qwen-32B, RLVR consistently raises CoT-Pass@K across K, supporting the claim that RLVR implicitly incentivizes correct reasoning paths.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "caea30a8-7a4d-46e7-9dc9-41a15a3c6a38",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Pass@K",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-41ac-7180-a81d-17318a1f70ff",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.2,
        "w": 0.65,
        "x": 0.02,
        "y": 0.22
      },
      "kind": "paragraph",
      "text": "RL with Verifiable Rewards (RLVR) has driven recent progress (OpenAI o1, DeepSeek-R1) by training on answers that can be automatically checked: math scores, program tests, or exact matches. However, two recent studies disagree on what RLVR actually adds. One argues it mostly reshuffles sampling without creating new reasoning; the other shows gains once you score the reasoning chains themselves rather than just final answers. Together they map where RLVR helps today and where it stalls.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "08d7cafb-3a3e-412c-b4e9-22b756815c2a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.6,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "RL from verifiable rewards: promising, but the evidence cuts both ways",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "9017739d-a8cc-4217-a20f-cfde10d9331e",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Analytical method",
      "slug": "analytical-method",
      "agent": null,
      "layer": "slide",
      "matchId": "6d9213c6-094c-4d0e-965f-1b0c4ddf4ccd",
      "evidence": "diagram/process: Conceptual diagram of RLVR reasoning paths and performance charts",
      "confidence": 0.5
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 189,
      "from": 12,
      "beatId": "019dd95a-0682-776c-8e35-9f380673831f",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Facts (What)",
      "beatSlug": "triple-take-the-facts-what",
      "evidence": "Sections 1-2 lay out research findings and industry data with charts and case studies.",
      "position": 1,
      "confidence": 70,
      "parentBeatName": "Setup",
      "parentBeatSlug": "setup"
    },
    {
      "to": 89,
      "from": 12,
      "beatId": "019dd95a-0682-776c-8e35-ac68557d958b",
      "arcName": "The Onion",
      "arcSlug": "onion",
      "beatName": "First Layer",
      "beatSlug": "onion-first-layer",
      "evidence": "Research section unpacks technical layer beneath the headlines.",
      "position": 2,
      "confidence": 45,
      "parentBeatName": "Development",
      "parentBeatSlug": "development"
    }
  ],
  "loops": [],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}