Perplexity says the agent behind its Computer product now studies its own failures, not only its successes. The company published details of a training method built to close a specific gap in how AI labs typically teach these systems, and it published the messy, mixed results alongside the method.
The standard technique, called rejection sampling fine-tuning, throws out any session that ended badly and trains the model only on the sessions that worked. Perplexity argues that approach has two problems: it can reward a model for sloppy steps buried inside an otherwise successful session, and it discards the exact moments where a model reveals how it goes wrong.
Perplexity’s name for its fix is hint-guided self-distillation, and it layers a correction step on top of that imitation. When a turn is judged to be an error, whether the session ultimately succeeded or failed, the system generates a short hint: a correction built only from information the model already had at the moment of the mistake, so it cannot cheat by using hindsight. The same model then runs twice on the identical recorded session, once seeing that hint and once without it, and the run without the hint is nudged toward matching what its hint-equipped twin produced.
Locating those error turns takes two separate methods. User complaints get traced backward through the session by an annotation pipeline, because Perplexity found that the turn right before a complaint is only sometimes the actual source of it: an earlier turn is to blame about half the time instead, so three AI judges review each case and at least two of them have to agree before the pipeline flags which turn caused the problem. Tool-use errors get different treatment, with deterministic rules catching known failure patterns and model judgment handling the ambiguous cases. Perplexity says it strips out sessions containing personal information and any session from a user who opted out.
Perplexity ran this test on the unmodified base model before deploying anything, to confirm hints actually help. Two small samples of 40 turns each isolated different kinds of mistakes: when the right fix was already stated earlier in the conversation, adding a hint pushed the share rated fixed or on track up from 40.0 percent to 75.0 percent. When the fix instead had to be inferred rather than read directly, the gain was larger, climbing from 32.5 percent up to 80.0 percent. A larger sample of 985 tool-error turns showed the same pattern at scale: hints lifted the rate of avoiding a repeat failure from 75.1 up to 93.7 percent. Separately, the share of turns where the model actually took the corrected fix rose from 60.6 to 82.3 percent.
Results after full training are messier, and Perplexity says so itself. Offline, the tool-error rate dropped from 2.79 percent on the stock GLM 5.2 model to 1.35 percent for a checkpoint trained with rejection sampling alone, and down further to 0.87 percent once the correction method was layered in. The company is upfront that this comparison is not clean: each of those three checkpoints learned from a different training set, and broader benchmarks of task success moved in mixed directions, so a lower error rate on tools did not translate cleanly into the model finishing more tasks.
Online results came from A/B tests with roughly 100,000 users in each group. The earliest checkpoint posted a 2.82 percent tool-failure rate against the stock model’s 2.94 percent, a gap Perplexity says is not statistically meaningful. The one number Perplexity stands behind is narrower: pitting two of its own trained checkpoints against each other, it saw tool-call failures drop from 2.24 percent down to 1.77 percent, a relative reduction of 21.2 percent that the company calls statistically significant. User dissatisfaction held nearly flat, 2.58 percent against 2.54 percent for the stock model, another difference Perplexity says also falls short of significance.
Perplexity is explicit that none of this is a benchmark win against a rival model, or even against its own unmodified system. The one clean result is a single internal checkpoint beating another internal checkpoint on a narrow metric, and the company frames its ambition around needing fewer corrections in the first place, not building a system that merely recovers well once it has already gone wrong. For any team building agents that call external tools, the percentage itself matters less, since it comes from Perplexity’s own unreplicated tests. What matters is the mechanism: correcting a model at its actual failure points, checked against what it knew when it erred, appears to move a live production system more than retraining on successes ever did.
Perplexity Research published these findings on its own blog on September 21, 2026.
