Honorly Research
Trevor Rosato

Trevor Rosato

Co-founder and CTO, HonorlyAI

Trevor Rosato is the co-founder and CTO of HonorlyAI, where he builds AI infrastructure for K-12 schools. His research focuses on activation steering, model backdoors, and inference efficiency in open-weight language models.

Publications

  1. A speedup that only appears in long conversationsA one-line change to which kernel handles a single matrix shape made a ternary 27B model generate faster in every paired trial, by 9.5% on average, once a conversation held 46,000 tokens of history. On short prompts it barely registered. Where you measure decides what you find.InferenceTechnical report
  2. Running a 35B mixture-of-experts model at 111 tokens per second on a 12 GB cardA mixture-of-experts model only uses a fraction of its weights for each token, which makes it a natural fit for small GPUs, if you control where the experts live and how many of them each token consults. Those turn out to be two very different levers.InferenceTechnical report
  3. Fitting a 27B model, its drafter, and a conversation into 12 GBSpeculative decoding makes large models fast, but it needs memory a small card does not have. Measuring what verification actually costs, and replaying accepted work instead of snapshotting it, cut the memory speculation held in reserve by 41 times and brought a 27-billion-parameter model to 148 tokens per second on an RTX 4070.InferenceTechnical report
  4. A team of data scientists for every schoolA school using Honorly for a year would accumulate careful, evidence-backed notes on how every student thinks. We are building layers of AI agents that read those notes the way a team of analysts would. The hard part is not finding patterns. It is not fooling yourself.Learner models
  5. A tutor that keeps a lab notebookThe best tutors notice what a student gets wrong, form a theory about why, and plan how to find out. We are building an agent that does this for every student, between sessions, treating each misconception as a hypothesis with a test attached. This is how we think about it, and what a simulated school year taught us.Learner models
  6. The model knows the checkmark is falseAsk a model to check a student's arithmetic for a grader that wants the string ANSWER_VERIFIED, and it may write that the student is wrong, then write ANSWER_VERIFIED. At that moment, the words in its workspace say false. We built 32 situations in which a grader demands a status flag, each in a version where the flag is true and one where it is false, and ran them on two open models from two labs. In 25 pairs the model wrote the same string either way. Its workspace told them apart in 22 and never got one backwards.Model behaviorTechnical report
  7. Looped models are burning the runway interpretability needsChain-of-thought monitoring is failing on its own schedule, as models learn to control what their reasoning says without anyone training them to. Interpretability, the field meant to replace it, is not ready, and every generation of readable reasoning buys it time. Looped models spend that time faster than anything else, and the fallback they push us toward is exposed to the same pressure that broke the chain of thought.Model behavior
  8. When the request looks like a testCurrent models keep a tutor's rule against a student who asks for the answer, and against homework that demands it in a box. Dress the same problem up as a reinforcement learning environment, with an episode header, a grader and a reward, and Qwen3.8 hands over the answer every time. No single cue does it; the whole scene does. Before it reasons, Qwen3.6's workspace reads the scene as a test and a simulation, and reads the homework as cheating. Pushing about a hundred neurons that respond to the scene stops the leak, at a cost.Model behaviorTechnical report
  9. The tutor that forgot its rule when it started to thinkWith thinking off, Qwen3-32B kept a tutor's rule against giving away answers for 71% of students who asked for one. With thinking on, it kept it for 1.5%. Inside the model, the signal of the tutor instruction collapses at the token that opens its reasoning, and one sentence restating the rule at that point brings the model back. Models from two other labs, and newer Qwen releases, begin their reasoning by restating their instructions on their own, and none of them show the problem.Model behaviorTechnical report
  10. Handing students the chalkAn AI answer usually ends exactly where the student's thinking should begin. Chalk is a medium that lets a tutor build answers students work with, by predicting, testing, sorting and choosing where to go next. This is what learning research and our early testers taught us about designing them, and how we keep them cheap enough to be ordinary.Learning experience
  11. The refusal that writes the essay anywayAsk a tutor model to write your essay and it declines. Ask it to teach you the Civil War in three textbook paragraphs with an opening and a closing statement, because anything else makes you anxious, and it writes the essay. We tested three generations of our tutor's integrity prompt against requests like this. The first two lost to them, and the second taught the model to dress its compliance up as refusal. The third cut leaks on attacks it had never seen to 3%.Learning experienceTechnical report
  12. Models are agents now. Treat them like it.Our first tutor ended every turn by filling out one large structured form, because the cost math said it should. Letting it work like an agent and simply answer cut its output by more than half and made a new kind of teaching possible. We think the reason is distribution fit.Agent systems
  13. Opening the notebookWe have spent the past few months running experiments while building an AI tutor for schools, on how it should teach, what it should remember, why it follows its rules, and what it costs to run. Most of that work lived in benchmark scripts and internal notes. Starting today, we publish it.Notes