Two outcomes on the same tasks
Zero is the first number that catches my attention in Volundr’s new release. Blockway reports no repetition loops across 50 tasks tested in its own coding-agent harness, compared with 36 affected tasks in the Preview. Then comes the second number: solved tasks rise from 22 to 27. Reading both outcomes together reveals the shape of the improvement. Escaping repeated actions clears one obstacle on the path to a correct result. How the remaining failures occur becomes a separate research question.[1]
Repetition measures how a session proceeds; solved-task counts measure its outcome. An agent could finish an incorrect change quickly or reach a dead end without looping. The aggregate counts leave those possible explanations unresolved. I would therefore want to follow each previously looping task to its outcome in the new version. That pairing would help distinguish whether the additional solutions concentrate among examples whose repetition problem was removed.[1]
Training changes that travel together
Blockway offers a plausible mechanism: the new training includes long agent conversations and anti-repetition preference pairs drawn from the model’s own sessions. Continued pretraining advances at the same time, however. Preview came from step 1,000 of a 20,000-step run; v1 reaches its end. Length-preference training changes too. Comparing the releases therefore measures the combined result of these changes. Estimating the contribution of anti-repetition training alone would require comparisons that hold the other components fixed.[1]
Finishing pretraining or changing response lengths could also contribute to the additional solutions. These alternatives preserve the value of the reported behavioral change while narrowing the causal conclusion about which intervention produced which outcome. The 50-task subset runs in the company’s Codeway harness, using each model’s default sampling. Identical limits make the comparison useful. Different tasks and an independent team could test how far the change extends beyond the selected examples.[1]
The comparison on scientific questions
The gains also move in different directions across outcomes. On GPQA Diamond’s scientific-reasoning questions, the table lists 84.3 for Preview and 80.3 for v1. Each is that model’s highest result across 4 runs. Selecting the best attempt does not directly estimate the expected result of a typical attempt. Applying that rule to both models still leaves a single high score as an uncertain basis for a broad capability claim. This comparison is especially useful when carrying an improvement on coding tasks over to scientific question answering.[1]
For me, the release’s useful contribution is making a targeted change in repetition visible alongside task completion. A more informative comparison would pair the old and new outcomes on identical tasks, isolate anti-repetition training from other changes, and publish the distribution across all runs. That would address three distinct questions: which failure was removed, which work was actually finished, and how consistently the difference appears. A researcher using the model gains confidence when those answers strengthen together.[1]