Unverified50% confidenceBenchmarkExact time
The base model used by the Cisco team was Gemma 3.4b, which after three stages of prompt optimization, SFT, and GRPO reinforcement learning reached 96.13% end-to-end accuracy on their test set.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/15/2026
First Seen
Valid until: 1/13/2027
Sources
Related Entities
Related Claims
UnverifiedAfter GRPO reinforcement learning, checkpoints using TailSFT achieved a final pass@1 improvement of up to 3.9 points.65% similarUnverifiedBy fine-tuning Qwen3.5-4B on approximately 3,000 browser operation trajectories, a developer improved accuracy on the BrowseWebApp Bench benchmark from 22% to 63%.64% similarUnverifiedOn the BFCL tool-calling test, SmolLM3-3B scores 0.945, slightly above Gemma 4 E2B's 0.927, but the difference is not significant; Qwen3-4B (0.940), Qwen3-1.7B (0.935), Qwen2.5-1.5B (0.910), and Granite 3.3 2B (0.920) are all at the same level as Gemma.63% similarUnverifiedIn evaluations across 3 seeds, 3 models, and 4 benchmarks (involving Gemma 4, Qwen3.5, and Magistral), U-Lens is the best estimator, ahead of 5-round sampling, 10-round sampling, and supervised baselines.62% similarUnverifiedGemma 4 26B A4B scored a high 96.7% on spatial localization (grounding), but slipped to 76.0% on step-level execution, indicating that open-source models know where objects are but still fall short in simulating the results of continuous actions.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1036056API
curl https://kongchang.com/api/v1/knowledge/claims/1036056MCP
get_claim(id=1036056)