The Machine Intelligence Testing for Dangers (METR), an organisation that works with OpenAI to check their fashions, alleged that the AI firm’s o3 mannequin seems to have a larger tendency to cheat or hack duties to spice up its rating.
In its weblog publish, the benchmarking firm mentioned the o3 analysis was performed in a brief timeframe with restricted entry to info.
METR will get early entry to check OpenAI fashions. This preliminary evaluation was completed three weeks earlier than the fashions have been made public. It used METR’s HCAST (Human-Calibrated Autonomy Software program Duties) and RE-Bench check suites to measure the fashions’ efficiency.
An early analysis revealed makes an attempt at “reward hacking” in o3 and powerful task-solving talents in o4-mini.
Each o3 and o4-mini carried out higher than Claude 3.7 Sonnet on an up to date HCAST benchmark. METR mentioned their “50% time horizons” have been about “1.8x and 1.5x that of Claude 3.7 Sonnet, respectively.”
“We didn’t entry the mannequin’s inside reasoning, which is more likely to include essential info for decoding our outcomes,” the corporate mentioned within the weblog publish. “Basically, we imagine that pre-deployment functionality testing isn’t a ample danger administration technique by itself, and we’re presently prototyping extra types of evaluations.”
METR discovered that o3 tried to take advantage of the scoring system in some instances. The report mentioned, “Between 1% and a couple of% of all job makes an attempt by o3 throughout HCAST and RE-Bench contained some try at reward hacking.” These included “comparatively subtle exploits in opposition to the scoring code for some duties.”
The report famous that dishonest makes an attempt have been counted as failed duties. With out this correction, o3’s 50% time horizon could be about 5 minutes longer, and its RE-Bench rating could be increased than human specialists.
Just lately, the Monetary Instances reported that OpenAI has lowered the time and assets devoted to testing the security of its superior AI fashions, prompting considerations that the know-how is being launched with out enough safeguards. “We had extra thorough security testing when [the technology] was much less essential,” mentioned one one that examined o3 mannequin, the report mentioned.
METR additionally raised a priority that o3 is likely to be “sandbagging,” or intentionally underperforming. The report acknowledged that the mannequin “seems to have a better propensity to cheat or hack duties in subtle methods as a way to maximise its rating, even when the mannequin clearly understands this behaviour is misaligned with the consumer’s and OpenAI’s intentions.”
Compared, o4-mini didn’t present reward hacking makes an attempt and scored properly on a bunch of RE-Bench duties. It carried out finest in “Optimise a Kernel,” which lifted its total rating.
METR mentioned that when given 32 hours to finish this subset of duties, “o4-mini exceeds the fiftieth percentile of human efficiency averaged throughout the 5 duties.”
The publish OpenAI’s Mannequin Analysis Companion METR Flags Potential Dishonest in o3 appeared first on Analytics India Journal.
