Datasets:
Gemma 4 Pareto frontier, open model grader, more languages?
Hello,
I ran this benchmark on Gemma 4 31B Thinking with both sonnet 4.6 and gemma 4 as graders and found that Gemma 4 (sonnet grader) achieves 80.5% at cost/task of $0.00062 which is on the Pareto frontier of your evaluation. I believe it'd be meaningful to evaluate many of the newer open models such as Qwen3.5/3.8 or Deepseek V4-Flash 0731 so that the comparison is up-to-date. These models provide a much cheaper and accessible alternative for students/educators to integrate into their workflows. I can share the full evaluation results if requested. The evaluations were performed using vLLM-served Gemma 4 31B Thinking at full precision.
Interestingly, the same evaluation evaluated by Gemma 4 as the grader yielded an 8pp difference (88%); this would be important to investigate as the Sonnet grader seems to be much harsher than Gemma 4 and the grader bias might be quite high for some reason.
For accessibility purposes, I hope the benchmark can also be run on an open-weight grader (Gemma 4, Qwen, etc.) so that the benchmark evaluation is not gatekept behind a closed-weights provider.
I was also wondering why the benchmark is only restricted to English as a Second Language learning. All the data points are on English-learning and there is no multi-lingual sample. Are there plans to add samples involving learning of another language (Any to Any) or English-learning in another language (Any-English)? I believe that the addition of such samples would be very helpful as language models are often used by students in their native languages to get feedback on english inputs, and such distribution shifts in the input modality often lead to different performance.
Thank you for the benchmark.
Hi,
Thank you so much for taking the time to run this and for the detailed writeup. This is exactly the kind of scrutiny we were hoping the benchmark would get. Also, apologies for the slow reply.
To address the points you made:
RE Updating results with the latest models: We will be launching a live leaderboard on our website in the coming weeks, incorporating new model evaluations as they become available (we will be sure to include the latest Qwens and Deepseeks as you point out). This results data will also be made available on EvalCards via the EEE datastore.
RE Gemma 4 data: Yes, we'd love to see the full evaluation results, please do share them - it might be worth checking out EEE datastore and Evalcards above as a means to share them? (we have only aggregates there currently, but intend to share our task-level instance data in the coming days). As you say, 80.5% at $0.00062/task would indeed sit on the Pareto frontier; we'll be sure to include it in our public results once we've had a chance to verify the run details.
RE Open-weight grader: Good catch, this is something we did cover briefly in our paper, but we will be revising this by adding further details and clarifying remarks shortly: Appendix H of the current (v2) paper covers judge performance and explains why we chose Sonnet over other open models like DeepSeek and Kimi; the upcoming v3 revision will explicitly call out DeepSeek v2 as a solid open-source backup judge. Our working hypothesis is that Gemma 4 is likely too small to be a reliable+performant judge, though we'll be able to confirm that properly in a dedicated follow-up paper doing a deeper dive across more open-weight judges both large and small (see Appendix A). Would you be willing to share the full evaluation results with Gemma 4 so we can look into its performance against our practitioner validation data?
RE Multilingual: Multilingual coverage is very much on our roadmap (see Appendix A), with prioritisation in world languages to begin (e.g. Spanish and Mandarin Chinese). We focused initially on English-learning as the target language to establish a baseline, but we note that we do have a diverse range of first (L1) languages in the tasks (see Appendix F.4). You make a great point on responses often being in the learner's L1; in some cases we certainly did account for this in our rubrics and reference answers, but we'll take an action to quantify how often our reference answers respond in this way. This multi-lingual extension to other target languages will take time and collaboration; we're very much open to collaboration on this.
Thanks again, this is genuinely useful feedback, and we're grateful you took the time to reach out.
Best wishes,
L2-Bench team