When RLHF Fails: A Mechanistic Taxonomy of Reward Hacking, Collapse, and Evaluator Gaming

When RLHF Fails: A Mechanistic Taxonomy of Reward Hacking, Collapse, and Evaluator Gaming | Signal Canvas | ScienceToStartup

Page Freshness

Signal Canvas proof surface

Canonical route: /signal-canvas/when-rlhf-fails-a-mechanistic-taxonomy-of-reward-hacking-collapse-and-evaluator-gaming

ready

Proof freshness: fresh
Proof status: unverified
Display score: 7/10
Last proof check: 2026-06-03
Score updated: 2026-06-03
Score fresh until: 2026-07-03
References: 0
Source count: 3
Coverage: 50%

Page-specific freshness sourced from this paper's evidence receipt and score bundle.

Agent Handoff

Canonical ID when-rlhf-fails-a-mechanistic-taxonomy-of-reward-hacking-collapse-and-evaluator-gaming | Route /signal-canvas/when-rlhf-fails-a-mechanistic-taxonomy-of-reward-hacking-collapse-and-evaluator-gaming

REST example

curl https://sciencetostartup.com/api/v1/agent-handoff/signal-canvas/when-rlhf-fails-a-mechanistic-taxonomy-of-reward-hacking-collapse-and-evaluator-gaming

MCP example

{
  "tool": "search_signal_canvas",
  "arguments": {
    "mode": "paper",
    "paper_ref": "when-rlhf-fails-a-mechanistic-taxonomy-of-reward-hacking-collapse-and-evaluator-gaming",
    "query_text": "Summarize When RLHF Fails: A Mechanistic Taxonomy of Reward Hacking, Collapse, and Evaluator Gaming"
  }
}

source_context

{
  "surface": "signal_canvas",
  "mode": "paper",
  "query": "When RLHF Fails: A Mechanistic Taxonomy of Reward Hacking, Collapse, and Evaluator Gaming",
  "normalized_query": "2606.03238",
  "route": "/signal-canvas/when-rlhf-fails-a-mechanistic-taxonomy-of-reward-hacking-collapse-and-evaluator-gaming",
  "paper_ref": "when-rlhf-fails-a-mechanistic-taxonomy-of-reward-hacking-collapse-and-evaluator-gaming",
  "topic_slug": null,
  "benchmark_ref": null,
  "dataset_ref": null
}

Evidence Receipt

Route status: building

Claims: 1

References: Pending verification

Proof: Verification pending

Freshness state: computing

Source paper: When RLHF Fails: A Mechanistic Taxonomy of Reward Hacking, Collapse, and Evaluator Gaming

PDF: https://arxiv.org/pdf/2606.03238v1

Source count: 3

Coverage: 50%

Last proof check: 2026-06-03T20:43:31.790Z

Signal Canvas receipt window

Watch and verify: When RLHF Fails: A Mechanistic Taxonomy of Reward Hacking, Collapse, and Evaluator Gaming

/buildability/when-rlhf-fails-a-mechanistic-taxonomy-of-reward-hacking-collapse-and-evaluator-gaming

Watchwatch

Subject: When RLHF Fails: A Mechanistic Taxonomy of Reward Hacking, Collapse, and Evaluator Gaming

Verdict

Watch

Verdict is Watch because viability or proof quality is intermediate and should be re-evaluated before execution.

When RLHF Fails: A Mechanistic Taxonomy of Reward Hacking, Collapse, and Evaluator Gaming

Use Signal Canvas as the narrative proof surface

Use this Signal Canvas via API or MCP

Signal Canvas proof surface