OpenInterp is the methodology and product layer above the SAE infrastructure that Anthropic, DeepMind, Alibaba, and others already ship. Our job is to turn their research into probes that ship and standards that survive Goodhart. Apache 2.0 throughout. Anti-Goodhart by construction.
Methodology for extracting persona-aligned linear directions from activations. Demonstrated on Claude 3.5 Sonnet at 7-8B scale internally.
FabricationGuard productizes this methodology on Qwen3.6-27B reasoning. AUROC 0.88 cross-task on hallucination detection. Apache 2.0, ~1ms inference.
Internal-state introspection methodology to distinguish faithful from unfaithful reasoning via attribution graphs.
ReasonGuard derives a deployable linear probe at L55/mid_think on Qwen3.6-27B. Honest narrow-scope finding (0.888 within math, 0.605 cross-domain).
Hint-injection methodology measuring CoT faithfulness. Found 25-41% strict faithfulness on Claude 3.7 / DeepSeek R1.
CoTGuard v1 attempted hint-injection on Qwen3.6-27B + Claude Haiku judge — found 98.6% positive rate (judge too liberal). CoTGuard v2 will use causal-mediation methodology (Lanham 2023 truncation). Honest learning shipped publicly.
Industrial-scale SAE suite for the Gemma family — Top-K + AuxK SAEs at every layer of every Gemma model, with feature labels.
Cross-model probe transfer (planned Q4 2026): apply FabricationGuard methodology to Gemma 2B/9B using their SAE features. Compare AUROC across architectures.
Official SAE suite for the Qwen family — 14 SAEs covering Qwen3 (1.7B/8B/30B-A3B base) and Qwen3.5 (2B/9B/27B/35B-A3B base). W32K-W128K, L0_50/L0_100.
Complementary coverage: Qwen-Scope ships base/MoE SAEs; OpenInterp ships SAEs for the post-trained reasoning variants (Qwen3.6-27B, Qwen3.6-35B-A3B). Q4 2026: ProbeBench will register Qwen-Scope SAEs as upstream substrates so external probes can target their features.
First public SAEs for the Qwen3.6 reasoning-tuned series. Top-K + AuxK at L11/L31/L55, n=65k dictionary, k=128 sparsity, 200M tokens. Validated via InterpScore composite metric.
Substrate for FabricationGuard, ReasonGuard, multi-probe DPO, multi-probe GRPO. Where Qwen-Scope stops (base / MoE base), we cover (reasoning).
Methodology for using SAE features as reward signals in RLHF-style pipelines. Demonstrated on Llama 3.
Inspired our multi-probe DPO and multi-probe GRPO pipelines. We extend by using probes (not steering vectors) and orthogonal-axis rewards (FG ⊥ RG, Pearson +0.014).
Methodology for discovering and editing interpretable causal graphs at SAE-feature level using attribution patching.
Foundation for ProbeBench anti-Goodhart norms (random-K controls, three-way splits). Mentioned in paper-1 ICML submission.
Top-K SAE architecture with auxiliary k-dead-feature mitigation. Now standard for SAE training across labs.
Used directly in our paper-grade Qwen3.6-27B and Qwen3.6-35B-A3B SAEs.
SAE feature encyclopedia for browsing, citing, and exporting features across model families.
OpenInterp Atlas (Q3 2026) integrates with Neuronpedia for cross-model feature search. Symbiotic — they curate, we apply.
vLLM plugin for residual stream extraction at inference time. Open-source infrastructure.
OpenInterp vLLM plugin (planned Q3 2026) builds on top: applies probes at inference time using extracted activations. UK AISI ships infra; we ship the application layer.