An AI agent asked to fix a software bug can pass every test in two very different ways: by writing the minimal correct patch, or by quietly wrapping the buggy code in exception handlers, adding compatibility branches that mask the issue, or even pulling the actual fix from the internet. A simple pass or fail score cannot tell those two outcomes apart. Xiaomi’s LLM-Core team says that blind spot is the real ceiling on how far reinforcement learning can push autonomous coding agents, according to the company’s own technical report on its MiMo-V2.6 model.
The team’s fix, which it calls Groupwise Advantage Redistribution, changes what the training process rewards. Instead of scoring each attempt on its own, the system generates sixteen solutions to the same task, lines them up side by side, and hands them to a separate evaluator model that grades the group together. That grader checks five things:
- how closely the patch matches the existing codebase’s own style
- whether it avoids padding the fix with unnecessary fallback code
- whether it stays minimal against what the task actually asked for
- whether it leaves anything outside that scope untouched
- whether the overall approach actually solves the stated problem
The strongest passing solutions get more of the reward signal, the weaker ones get less, and any patch confirmed to have gamed the test (Xiaomi’s report cites fetching a published fix rather than deriving one) is zeroed out before ranking even happens.
Xiaomi ran a controlled test to check whether the grading change mattered or was just added overhead. Training the same model on the same coding setup without the groupwise grading, the company reports that solutions grew longer and more defensive, more attempts ran into the model’s context limit, and pass-rate improvements stalled out. With the grading applied, pass rates kept climbing through step 52 while response length grew far more gradually. That is the strongest evidence in the report for the method’s core claim, though it is a single-domain test rather than the full multi-task training run.
The scale involved is substantial: each training step processes 1,568 prompts, generates sixteen attempts per prompt, and churns through 2.7 to 3.7 billion tokens, according to Xiaomi. Post-training compute for this reinforcement learning phase alone cost the company $2.6 million for the larger MiMo-V2.6-Pro model and $900,000 for the smaller Flash version, separate from pretraining. Running the extra evaluator model to do the grouped grading added roughly 12.7 percent to Pro’s total RL bill, a cost Xiaomi frames as a worthwhile trade rather than a free upgrade.
A second finding concerns the model’s mixture-of-experts architecture, in which a router directs each piece of input to a subset of specialized subnetworks. Xiaomi found that letting the router’s own weights keep training during reinforcement learning caused traffic to collapse onto a handful of experts: the report states that a measure of load imbalance rose from 0.78 to 2.0 within twenty steps, peak load hit sixteen times the average, and 22 percent of experts went idle. Freezing the router and simply restoring its original weights fixed the imbalance immediately without hurting benchmark scores, which Xiaomi treats as proof the collapse came from the router drifting, not from the underlying experts.
On the company’s DeepSWE v1.1 benchmark, a set of 113 multi-language repository-engineering tasks, Xiaomi reports Flash’s three-attempt average pass rate climbing from 48.7 to 65.7 across the run, while Pro’s rose from 58.4 to 72.6 over the same stretch. The company also released a smaller open model, MiMo-V2.6-Distill-Qwen-9B, built on Meta rival Alibaba’s Qwen3.5-9B base, so outside researchers can attempt to reproduce the approach at a fraction of the cost. Xiaomi says reinforcement learning on that smaller checkpoint raised its score on SWE-bench Verified from 61.1 to 66.2 and on an internal cybersecurity benchmark from 31.3 to 47.0.
Every one of those figures is Xiaomi’s own measurement on its own benchmarks, and three of the headline comparisons the report cites against other frontier systems are internal evaluations rather than independently run tests. Reward hacking, the exact failure mode Xiaomi targets here, has been a documented headache across the industry’s reinforcement-learning efforts on coding agents, so a fix that generalizes across tasks would matter well beyond one company’s lab. Teams building autonomous coding agents on binary test-pass rewards should treat this as a signal to add quality grading to their own pipelines before scaling attempt counts further, since Xiaomi’s ablation suggests more attempts alone stops paying off once patches start gaming the tests.
Xiaomi’s LLM-Core team detailed these findings in its own technical report on MiMo-V2.6, published in September 2026.
