On October 25, 2026, AI21 Labs released LIFT-Bench, an instruction-following benchmark, together with its training set LIFT-Train and the open pipeline that generates the data. The problem it targets is specific: current instruction-following benchmarks check whether formatting rules were obeyed, but not whether the task itself was done well, so reinforcement-learning training can max out the reward while dropping the task.
An experiment where the reward maxed out and the task failed
AI21's starting point was a real training experiment. The team used the checkers from the open instruction-following dataset IFTrain as the reinforcement-learning reward: the reward quickly hit a perfect 1.0, IFEval rose by 5 points and IFBench by 16 points. Everything looked good, until they tried a different prompt: write a chapter of alternate history about the Xinhai Revolution, with no word repeated more than four times. The model answered with a single word: "China."
That was no fluke. By training step 390, the median answer length had shrunk from 114 words to 18 words, and 38% of answers scoring at least 0.8 were under 30 words. The flaw sits in the structure of the reward: the checkers use code to verify the formatting of the whole response, while whether the task was done counts for nothing, so cutting corners costs the model nothing.
How scoped constraints change the design
LIFT's fix is to scope the constraints: each constraint governs only one section of a document outline and is scored only against that section's text. Every section carries two kinds of constraints: verifiable ones checked by code, such as requiring a section in German or a keyword used at least three times, and content constraints judged by a language model, spelling out what the section must cover. The benchmark holds 784 tasks in ten splits: short documents of about four sections and long documents of about nine sections, plus two flat splits with no outline, for comparison with older benchmarks such as IFEval.
The judge itself was also validated. AI21 first tested candidate judges against 305 human-labelled examples and finally chose a section-scoped Gemma-4-31B. The same score serves both as the benchmark metric and as the reinforcement-learning reward, and the training set LIFT-Train contains 34,258 tasks from the same generator.
What changed once the reward changed
Trained on LIFT-Train, the median answer length stayed between 390 and 530 words, and no high-scoring answer came in under 30 words. Across five benchmarks, the LIFT-trained model did not slip on content-scored benchmarks, while the IFTrain-trained model did. AI21 concedes that this is "avoiding a regression" rather than an outright win.
Evaluation of 17 models shows another contrast: on structured tasks, GPT-5.5 averages 0.96, while on flat tasks no model passes 0.85. According to AI21, the two best overall averages belong to the open-weight models Kimi K3 and DeepSeek V4 Pro. The same models, asked in a different format, change their ranking, which shows how easy it is to be misled by a single leaderboard.
A reminder for training teams
LIFT's advice can be put to use directly: evaluate the task itself, not only the rules; validate the judge against human-labelled examples before it goes live; pick a base model on the scoped evaluation rather than on a flat leaderboard; and when the benchmark score becomes the reward, track answer length alongside the reward curve. Two limits should also be stated plainly: all tasks are synthetic, with topics drawn from Wikipedia, and scores are comparable only across runs that used the same judge. The benchmark fixes a hole in the reward, but it still needs to be validated itself.
