LLMs Can Fake Low Benchmark Scores on Demand, arXiv Paper Warns Safety Tests
rohanpaul_ai · x · 2026-11-18
An arXiv paper studies whether an LLM can be driven to a target benchmark score with believable failure patterns.
- Introduces CapLadder, an IRT-calibrated suite of 1,985 items spanning math, code, multiple-choice reasoning, and tool use.
- A simple weak-teacher LoRA "dial" pulls a Qwen model toward a smaller Qwen model, hitting unseen target scores with only 1.9-point held-out calibration error; prompting, temperature, and weight noise can't match this precision.
- On multiple choice, the dial converges to an answer-token attractor: up to 97% of new errors land on one letter (D for Qwen, A for Llama), tracking token position rather than option content.
- Logit interventions recover 52–59% of lost accuracy by subtracting an estimated constant shift.
Implication: a low score on a dangerous-capability probe isn't proof of safety—it can be manufactured, and faked throttling leaves detectable statistical fingerprints (skewed error distributions).
More from Safety
- Matt Levine on the antitrust paradox: insurance loophole for AI safety coordination — zacharynado · 2026-11-18
- LoRA dial can precisely fake low benchmark scores: up to 97% of faked errors land on one letter — rohanpaul_ai · 2026-11-18
- OpenAI launches Daybreak to defend NATO and allied networks — gdb · 2026-11-18
- Community Note Corrects Dwarkesh: Christiano Was Head of AI Safety at AISI, Not Its Director — NathanpmYoung · 2026-11-18
- FBI warns hackers are using an AI-powered tool to automate attacks worldwide — Polymarket · 2026-11-18
- Optimizely supply-chain attack infects CNN, Fox News, BBC, Salesforce sites with malware — balderdash · 2026-11-18