Expired50% confidenceBenchmarkExact time
On the 78 tasks of the MMEB-V2 benchmark, PILOT outperforms the main baselines at both the 2B and 7B parameter scales.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/26/2026
First Seen
This claim has expired. It is kept for the record and no longer describes the current state.
Valid until: 11/9/2026(expired)
Sources
Related Entities
Related Claims
UnverifiedPILOT improves on all 12 task groups of MMEB-V2, raising the overall score by 3.2 and 3.3 points respectively (2B and 7B scales).67% similarUnverified在HumanEval、MBPP、GSM8K等基准测试上,Gemma 7B的表现显著优于同参数量级的LLaMA 2 7B和Mistral 7B65% similarUnverifiedIn the MoE architecture, the Mixtral 8x7B model has a total of 46.7B parameters, but only approximately 12.9B parameters (2 experts) are activated per token, making inference costs close to those of a 13B dense model while achieving performance comparable to the 70B level.63% similarUnverified在全部 38 项基准中,Aplomb 1 在 8 项上拿到 5.3B 以内模型最高分,包括 MMLU-Pro、GPQA Diamond 和 BBH63% similarVerifiedThe larger the number of model parameters, the greater the robustness to quantization, allowing 70B-level models to maintain performance close to FP16 after 4-bit quantization.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1098732API
curl https://kongchang.com/api/v1/knowledge/claims/1098732MCP
get_claim(id=1098732)