A frozen model, a growing archive of edits, and a 17-point average jump across four benchmarks
Salesforce AI Research's new DarwinX paper challenges the idea that frozen models cannot be substantially improved without full retraining. The work describes a Survival-of-the-Fittest pipeline that layers iterative edits from a growing archive onto a fixed base model, reporting an average gain of 17 points across four benchmarks. Reported scores include a 7.7-point rise on Terminal-Bench 2.1 and a leap from 43.5% to 93.0% on WebArena-Infinity. For builders, the signal is that curated post-training edits may deliver large benchmark jumps at lower compute than wholesale fine-tuning. The packet does not specify edit types, baseline model size, or benchmark definitions, so real-world transfer beyond these metrics remains unverified.
A frozen model, a growing archive of edits, and a 17-point average jump across four benchmarks
Salesforce AI Research just published a paper that pushes back on that assumption in a very direct way. The paper is called DarwinX, and the results are hard to wave away: a 7.7 point jump on Terminal-Bench 2.1, a jump from 43.5% to 93.0% on WebArena-Infinity.
Key takeaway
DarwinX shows that a frozen base model plus a growing archive of curated edits can lift benchmark scores sharply without full retraining.
What happened
Salesforce AI Research published a paper called DarwinX that pushes back on assumptions about improving frozen models, according to reporting summarized in the evidence packet.
The packet cites reported gains of a 7.7-point jump on Terminal-Bench 2.1, a move from 43.5% to 93.0% on WebArena-Infinity, and a 17-point average increase across four benchmarks using a growing archive of edits on a frozen base model.
Evidence
Salesforce AI Research published a paper called DarwinX with large reported benchmark gains.
Towards AI · attributed
Salesforce AI Research just published a paper that pushes back on that assumption in a very direct way. The paper is called DarwinX, and the results are hard to wave away: a 7.7 point jump on Terminal-Bench 2.1, a jump from 43.5% to 93.0% on WebArena-Infinity.
A frozen base model was improved via a growing archive of edits, yielding a 17-point average jump across four benchmarks.
Towards AI · attributed
A frozen base model was improved by applying edits from a growing archive, resulting in a 17-point average jump across four benchmarks.
The Survival-of-the-Fittest approach uses iterative fine-tuning on curated archives without full model retraining.
Towards AI · attributed
The 'Survival of the Fittest' approach suggests that iterative fine-tuning on curated, growing archives can yield significant benchmark gains without requiring full model retraining.
Why it matters
If edit-archive post-training reliably scales on frozen weights, teams could customize large models with less compute than full fine-tuning or retraining from scratch.
Limits and uncertainties
The packet does not describe the nature of the edits, the baseline model size, or the four benchmarks used, leaving generalizability of the 17-point average jump unclear.
Practical implications
Operators evaluating post-training stacks should compare edit-archive pipelines like DarwinX against full fine-tuning for target agent benchmarks such as Terminal-Bench and WebArena-Infinity.
What to watch
Publication of DarwinX methodology details, including edit sources, archive curation rules, and full benchmark names beyond the cited Terminal-Bench 2.1 and WebArena-Infinity results.
