01 / 03
What can models do, where do they fail, and how reliably can we tell?
Benchmarks, robustness tests and evaluation protocols built to be audited rather than trusted: per-problem provenance, verifier-checkable answers, and an explicit limit on what a score supports.
NeurIPS 2026
A benchmark that ships its own audit: every problem carries 13-gram contamination-overlap evidence against public pretraining text. 4,027 problems drawn from 836 arXiv papers across 20 wireless subfields; frontier models cluster at 86.5–91.3%.
Scope A rerunnable audit producing per-problem evidence for one lexical channel — not a proof that the benchmark is clean.
ICLR 2026
A model can verify a function on its own and still fail once the specifications have to compose. The benchmark measures the distance between those two results.
ACL 2026 Findings
400+ SWE-level task instances built from real Ascend CANN repositories — multi-file, multi-language and execution-aware — on a live benchmarking paradigm that mitigates leakage.
TMLR 2026
A benchmark of 4,510 multiple-choice and 7,380 open-ended inspection questions on 410 industrial images, testing whether ~4B multimodal models can serve as on-device anomaly-detection assistants: the best, Qwen3-VL-4B, scores 88.3% on multiple choice against 82.3% for GPT-5 Nano, yet only 72.8% of its open-ended answers reach a passing 3 of 5.
ACL 2025 Findings
The first benchmark for LLM mathematical reasoning in wireless communications, and the work WirelessMathBench-XL extends.