New Frontier data and RL environments, off the shelf

Post-training Runs

We train frontier models on Surge datasets, then study what they learn: benchmark gains, transfer beyond the training distribution, and the behaviors that emerge.

What improved?

Which behaviors and benchmarks moved?

what transferred?

Did gains survive outside the training distribution?

what changed?

How did the model behave differently?

Agentic Coding
·
Kimi K2.7
Aug 2026

We trained a model on 1,700 coding tasks. It learned skills that generalized across every SWE eval.

Training on 1,700 expert-built coding tasks improved Kimi K2.7 across every external SWE eval we measured. The biggest behavioral changes showed up in the complexity of real-world software engineering: carrying messy specifications through, testing beyond the obvious path, preserving existing behavior, and building its own checks when the environment did not provide one.

SWE-Marathon
pass@1
+20.0pp
5.0% → 25.0%
Terminal-Bench 2.1
pass@1
+14.6pp
67.4% → 82.0%
DeepSWE
pass@1
+12.4pp
31.0% → 43.4%
Terminal-Bench 3
pass@1
+10.7pp
1.4% → 12.1%
SWE-Bench Pro
pass@1
+4.7pp
60.1% → 64.8%
Goal-Directed Execution
·
Qwen3.5-122B-A10B
Jul 2026

We trained a model on office work RL environments with no code. It got better at coding anyways.

We trained on long-horizon office tasks with no coding examples. SWE-Bench Pro still improved +5.8pp. Trajectory analysis pointed to a broader capability we call goal-directed execution: choosing useful subgoals, maintaining task state, staying faithful to the objective, and verifying completion.

SWE-Bench Pro
pass@1 · zero coding tasks in training
+5.8pp
External transfer
Complex Instruction Following
·
Qwen3-4B-Thinking
Jun 2026

We trained a model on single-turn ComplexConstraints. It also got better at multi-turn instruction following.

Every ComplexConstraints training example was single-turn, yet the largest gains appeared in multi-turn behaviors. Our hypothesis is that the underlying skill is shared: maintaining many interacting constraints without dropping them, whether they arrive in one dense prompt or accumulate over nine turns.

MultiChallenge
overall
+10.1pp
41.1% → 51.2%
AdvancedIF
pass@1
+8.4pp
28.2% → 36.6%
Instruction Retention
multi-turn dimension
+22.1pp
Largest behavioral gain
Tool Use Agents
·
Qwen3.5-122B-A10B
May 2026

We trained a model on long-horizon tool-use environments. It learned agentic behaviors that transferred across every tool eval.

We trained Qwen3.5-122B-A10B on realistic, multi-tool agent environments. The gains transferred to Toolathlon, τ²-Bench, and BFCL-V4. Trajectory analysis showed better task closure and more efficient tool use across unfamiliar evaluation environments.

Toolathlon
pass@1
+9.6pp
24.2% → 33.8%
τ²-Bench
pass@1
+5.3pp
54.8% → 60.1%
BFCL-V4
pass@1
+3.5pp
55.7% → 59.2%
Expert STEM
·
GLM-4.7
May 2026

We trained a model on expert STEM problems. The gains transferred across every scientific reasoning benchmark.

We trained GLM-4.7 on expert-authored STEM problems. It nearly doubled performance on held-out Expert STEM tasks, with gains also appearing on HLE, FrontierScience, PubMedQA, and DeepDive. The result suggests that difficult expert problem solving can teach reasoning patterns that remain useful across different scientific domains.

Expert STEM
pass@1 · held-out
+24.9pp
22.7% → 47.6%
HLE
pass@1 · text prompts
+6.8pp
26.8% → 33.6%
FrontierScience
pass@1
+9.3pp
29.1% → 38.4%
PubMedQA
pass@1
+13.7pp
66.7% → 80.4%
DeepDive
pass@1
+9.1pp
57.4% → 66.5%
Enterprise Agents
·
GLM-4.6
May 2026

We trained a model on customer-support RL environments. It learned tool-use skills that generalized beyond.

We trained GLM-4.6 inside CoreCraft, a high-fidelity customer-support environment. The gains transferred beyond CoreCraft to BFCL, τ²-Bench Retail, and Toolathlon. The result suggests that realistic enterprise workflows can teach general agent skills that survive changes in tools, tasks, and evaluation environments.

CoreCraft
task pass rate · held-out
+11.4pp
25.4% → 36.8%
BFCL Parallel
task pass rate
+4.5pp
91.0% → 95.5%
BFCL Simple
task pass rate
+2.0pp
91.5% → 93.5%
τ²-Bench Retail
task pass rate
+7.4pp
68.7% → 76.1%
Toolathlon
task pass rate
+6.8pp
18.8% → 25.6%

See what rich, high-quality datasets teach your models

All of these experiments use Surge datasets available off the shelf. Run the training data on your model, measure what changes, and see whether the capabilities transfer.

Train on the full dataset first. Pay only if it moves the metrics that matter.