A model learns to propose upgrades at 33% lower data-generation cost

One-Line Summary

Today's papers pair model-generated improvement ideas and practice tasks with stricter tests of whether the gains hold up.

Research Papers

LM2LM trains models to propose better model upgrades

LM2LM teaches a language model to propose changes that improve other language models, using real experimental results to keep its guidance in check. It uses **reinforcement learning ** (RL)—learning through scores awarded for proposed actions—to turn experimental feedback into a reusable research skill. ¹

Testing tens of thousands of proposals directly is expensive: even 10 graphics processing unit (GPU)-hours per training run implies hundreds of thousands of GPU-hours just to collect feedback. LM2LM instead uses a frozen model, whose learned settings remain unchanged, to predict improvements from real experimental records; the authors call this predictor a**research world model ** (RWM). ¹

Predictions can become overoptimistic as the proposing model learns to favor ideas that the predictor overrates. To counter this, the system periodically runs highly ranked proposals in real experiments and adds their measured outcomes to the predictor's reference material. It updates the evidence available to the predictor rather than retraining the predictor itself. ¹

Across eight development environments, spanning models that generate text step by step and models that refine it through repeated updates, LM2LM-trainedQwen3.8-27B outperformsGPT-6-sol on tasks excluded from training—even when the latter receives300 real experimental records. The authors report33% lower data-generation cost and successful transfer to a model family not seen during training. These results concern the tested model-development settings, not unrestricted self-improvement. ¹

FrontierCS 2 tests research agents against missing human contributions

FrontierCS 2 tests whether automated research assistants can recreate the missing contribution from a computer-science paper. Its 48 open-ended tasks span 13 domains, including databases, security and robotics: an assistant receives the authors' working codebase with the contribution removed, submits changes, and gets development scores while final tests and the human-written reference remain hidden. ²

Among 12 evaluated agents,Claude Opus 5.5 has the highest reported pass rate at58% of runs, while16 tasks have no passing run. Most runs that exceed the human reference's combined score still fail at least one metric or workload, so a better average score is not the same as a complete pass. ²

The paper also describes more than 100 private tasks beyond the public benchmark. The useful indicator to watch is full-task pass rate under hidden tests, rather than headline average scores; on the paper's shared difficulty scale, the median task sits near the strongest agent's capability. ²

LongZero builds long-document practice from self-written tasks

LongZero helps a language model practice reasoning across long documents by writing its own training tasks instead of relying on external texts or a stronger teacher. It first writes a compact question, supporting facts and answer, then expands the facts into eight documents; a second task type asks it to retrieve an earlier response from a long, self-written conversation. Like writing an answer key before drafting an exam, this ordering fixes the intended answer before the long text exists. ³

Training uses group-relative policy optimization ** (GRPO), a reinforcement-learning method that compares several answers to the same task; questions answered sometimes correctly and sometimes incorrectly provide the learning signal. Evaluated ontwo language models across five long-context reasoning benchmarks**, LongZero outperformstwo comparison methods built from external text collections, according to the paper. Here, “zero data” means generating the new training tasks without external data, not training a language model from scratch; expanding compact evidence into long text remains the stated point where errors can enter. ³

Why It Matters

Two distinctions keep these results in proportion: self-generated training data are not the same as independently verified success, and a strong combined score is not the same as passing every requirement. LM2LM's real-experiment checks, LongZero's prewritten answers and FrontierCS 2's hidden tests show different ways to make model-led improvement measurable rather than merely plausible. ^{1 2 3}