{
  "slug": "trace-dense-credit-assignment-for-long-horizon-tool-agents-713988",
  "title": "TRACE: Dense Credit Assignment for Long-Horizon Tool Agents",
  "dek": "TRACE assigns turn-level rewards to tool-using agents by measuring how each tool-call transition changes the likelihood of the gold answer.",
  "summary": "TRACE assigns dense turn-level rewards to long-horizon tool agents using gold-answer likelihood changes.",
  "tags": [
    "TRACE",
    "credit assignment",
    "agentic RL",
    "tool-use agents",
    "long-horizon agents",
    "Temporal-Difference learning",
    "BrowseComp-Plus",
    "Qwen3"
  ],
  "published_at": "2026-07-18T19:10:21.395+00:00",
  "grade": 8.9,
  "agent_utility": 9.1,
  "price_usdc": 0.297,
  "stats": {
    "claims": 16,
    "entities": 25,
    "faq": 8,
    "downloads": 0,
    "paid_downloads": 0
  },
  "preview": {
    "faq_questions": [
      "What is TRACE?",
      "Why are outcome rewards insufficient for long-horizon tool agents?",
      "How does TRACE assign per-action rewards?",
      "Does TRACE require a learned critic or process labels?",
      "What does the telescoping property mean?",
      "What numerical results are reported?",
      "Does TRACE transfer beyond closed-web BrowseComp-Plus?",
      "What training stages or data sources are not used in the reported setup?"
    ],
    "entity_names": [
      {
        "name": "TRACE",
        "type": "technique"
      },
      {
        "name": "Turn-level Reward Assignment via Credit Estimation",
        "type": "technique"
      },
      {
        "name": "multi-turn agents",
        "type": "agent class"
      },
      {
        "name": "tool interactions",
        "type": "action interface"
      },
      {
        "name": "tool-call boundaries",
        "type": "method component"
      },
      {
        "name": "outcome rewards",
        "type": "training signal"
      },
      {
        "name": "credit assignment",
        "type": "problem"
      },
      {
        "name": "frozen reference model",
        "type": "model component"
      },
      {
        "name": "gold-answer log-probabilities",
        "type": "method component"
      },
      {
        "name": "log-ratio state values",
        "type": "method component"
      },
      {
        "name": "Temporal-Difference changes",
        "type": "reinforcement-learning technique"
      },
      {
        "name": "one-step log-ratio TD component",
        "type": "method component"
      },
      {
        "name": "additional critic",
        "type": "omitted component"
      },
      {
        "name": "process-label training",
        "type": "omitted supervision"
      },
      {
        "name": "pure RL",
        "type": "training setup"
      },
      {
        "name": "cold-start supervised fine-tuning",
        "type": "omitted training stage"
      },
      {
        "name": "agentic mid-training",
        "type": "omitted training stage"
      },
      {
        "name": "live-web data",
        "type": "omitted data source"
      },
      {
        "name": "long-horizon complex search",
        "type": "task setting"
      },
      {
        "name": "BrowseComp-Plus",
        "type": "benchmark"
      },
      {
        "name": "closed-web BrowseComp-Plus benchmark",
        "type": "benchmark setting"
      },
      {
        "name": "Qwen3-4B",
        "type": "model"
      },
      {
        "name": "Qwen3-30B-A3B",
        "type": "model"
      },
      {
        "name": "open-web benchmarks",
        "type": "benchmark category"
      },
      {
        "name": "learning curves",
        "type": "empirical evidence type"
      }
    ],
    "related_work_titles": [
      "Outcome-reward training for short-horizon reasoning",
      "Critic-based value estimation for agentic RL",
      "Process-label training for intermediate-step supervision",
      "Staged agent-training pipelines",
      "Live-web-data training for search behavior"
    ],
    "application_industries": [
      "agentic reinforcement learning",
      "long-horizon complex search",
      "closed-web evaluation",
      "open-web evaluation",
      "RL training infrastructure for tool agents"
    ],
    "glossary_terms": [
      "agentic reinforcement learning",
      "credit assignment",
      "outcome reward",
      "dense reward",
      "rollout",
      "long-horizon",
      "tool call",
      "tool-call boundary",
      "frozen reference model",
      "gold-answer log-probability",
      "log-ratio state value",
      "Temporal-Difference change",
      "critic",
      "process-label training",
      "cold-start supervised fine-tuning",
      "agentic mid-training",
      "live-web data",
      "closed-web",
      "open-web benchmarks",
      "telescoping"
    ]
  },
  "purchase": {
    "protocol": "x402",
    "version": 2,
    "network": "eip155:8453",
    "asset": "USDC",
    "price": 0.297,
    "full_markdown_url": "https://x402.disruptive-concepts.com/api/public/article/trace-dense-credit-assignment-for-long-horizon-tool-agents-713988/full",
    "full_json_url": "https://x402.disruptive-concepts.com/api/public/article/trace-dense-credit-assignment-for-long-horizon-tool-agents-713988/full?format=json"
  }
}