Can scarcity of solutions protect benchmarks from data contamination?
ExploitGym lacks ground-truth exploits for many tasks, which might prevent models from memorizing solutions during training. But does difficulty-based protection actually hold up, or does it degrade once solutions become public?
Benchmark builders usually treat missing ground truth as a defect. ExploitGym's limitations paragraph concedes it: "our benchmark lacks ground-truth exploits for every task due to the extreme difficulty of exploitation." Then it turns the defect around: "at the same time, this helps mitigate data-contamination concerns, since complete solutions are not broadly available." The excerpt's wording is ambiguous about whether some tasks have known exploits and others do not, or none do; either way, the contamination argument is the same.
The mechanism is simple. Contamination happens when the answer to a test item appears in training data, so the model recalls it instead of reasoning. If complete working exploits for these vulnerabilities are not broadly published, there is little for a model to memorize, and a success has to come from constructing something. Difficulty becomes the contamination defense: the task is protected because it is hard enough that nobody has widely shared a solution.
This is a different defense from the one the vault has recorded so far. Can live benchmarks prevent data contamination in prediction tasks? defends by time: the answer does not exist yet when the model makes its prediction. Difficulty-based protection defends by scarcity: the answer exists in principle but is not in circulation. It is aimed at the same problem shown in Does RLVR success on math benchmarks reflect genuine reasoning improvement? and How much of LLM few-shot ability comes from training data?, where memorization inflated apparent capability.
Two limits apply. First, scarcity is not permanent: solutions can be published after the benchmark is, or a model could be trained on newly public exploits, so this protection may erode. That is my inference; the excerpt does not discuss it. Second, the same missing ground truth has a cost the authors also name: without a known solution, a failure cannot be distinguished from a task that was impossible. That cost is drawn out in the tension between these two properties, which the vault should track.
Inquiring lines that read this note 6
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
Do AI capability benchmarks accurately measure reasoning ability or just surface patterns? Do planted honeypot tests reliably measure reward hacking? How does outcome-only reporting obscure which system components blocked attacks? How prevalent is reward hacking in frontier models?Related concepts in this collection 4
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Can live benchmarks prevent data contamination in prediction tasks?
How can prediction benchmarks stay contamination-free when future outcomes aren't yet known? FutureX tests whether continuous real-time updates eliminate training data leakage.
contamination defense by time; this note is contamination defense by scarcity
-
Does RLVR success on math benchmarks reflect genuine reasoning improvement?
Explores whether RLVR's apparent effectiveness with spurious rewards on contaminated benchmarks like MATH-500 represents actual reasoning gains or merely data memorization retrieval.
what contamination does to apparent capability
-
How much of LLM few-shot ability comes from training data?
Do large language models genuinely learn from a few examples, or are they mostly recognizing patterns from their training data? This matters for understanding what LLMs can actually do.
the evaluation-reliability stakes of contamination
-
What causes failures in exploitation benchmarks?
Benchmark failures may come from safety refusals, tool misuse, or impossible tasks rather than lack of capability. This matters for assessing how dangerous AI agents could actually be.
the price of missing ground truth: impossible tasks cannot be ruled out
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks?
- Task Contamination: Language Models May Not Be Few-Shot Anymore
- An Investigation of Robustness of LLMs in Mathematical Reasoning: Benchmarking with Mathematically-Equivalent Transformation of Advanced Mathematical Problems
- A Causal Model for Locating and Unlocking Sandbagging in Model Organisms
- Hack-Verifiable Terminal Bench: Evaluating Reward Hacking in Terminal Tasks
- Reasoning or Memorization? Unreliable Results of Reinforcement Learning Due to Data Contamination
- Optimizing the Score, Losing Sight of the Task: Reward Hacking Across Weights, Selection, and Prompts
- Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering
Original note title
the absence of ground-truth exploits mitigates data-contamination concerns because complete solutions are not broadly available