NeurIPS 2026: a 120B model was no more fault-tolerant than a 20B one once MCP tools started failing
AspireAI community · arXiv (University of Luxembourg SNT)
MCP-Injector, a protocol-aware fault-injection middleware, mined 57,240 issue and pull-request threads across 1,774 open-source MCP repositories to build a fault model calibrated to what actually breaks in the wild, then crossed two gpt-oss backbone sizes (20B and 120B), three agent stacks (a custom multi-round plan-and-execute loop, the OpenAI Agents SDK, and LangChain) and three injection conditions. Under a permanent-heavy stress profile, task-completion scores fell 60 percent, from 8.08 to 3.23 out of 10. The headline is the resilience-ratio result: the 120B model retained no larger share of its fault-free score than the 20B (p = 0.96), and its completion rate under faults was if anything lower, 91.1 percent versus 95.6 percent, a 4.4 point difference with a 95 percent confidence interval of [-0.2, 9.3] that does not exclude zero. The authors' conclusion is that agent stacks differed more in how well they absorbed faults than the backbones did, so framework-level error handling matters more than model scale. The honest caveat: this is a synthetic fault model derived from issue-thread mining, not a live production incident study, and the confidence interval on the scale comparison is wide enough that the authors themselves frame the completion-rate gap as suggestive rather than established.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.