A faithful activation reader should notice when the thing asked about is absent from its activations and should know which of its statements the activations sup
Diagnosing a model's reasoning with an Activation Oracle (AO) requires reading trajectory-specific facts, not just problem-level gist. We test whether AO reader
Route B of this problem asks whether reader updates trained on near-miss description pairs reduce fabricated detail in Activation Oracle readouts relative to or
Activation Oracles (AOs) face a sharp specificity-faithfulness trade-off: three-layer supervised readers (`ours-3layers`) recover more target attributes than si
We compare Activation Oracle (AO) readers of Qwen3-8B with matched text readers and donor activations in two settings. On reasoning traces we count rollout-spec
Do activation-contrastive objectives make fine-tuned Activation Oracle (AO) readers report numbers more faithfully than plain fine-tuning? We test this on a que
What can an Activation Oracle reader say about a model's next step from one end-of-window activation, beyond the visible text? We test two Qwen3-8B readers (GRP
Answering entity-specific questions about multi-step `<think>` chains-of-thought (`CoT`) with a frozen Activation Oracle (AO) exposes a spatial-scale incompatib
PersonaQA asks an activation reader for attributes a fine-tuned model stores for 100 personas but its context does not state. The adapted Qwen3-8B answers 93%;
Activation Oracle (AO) readers score 9–17% on PersonaQA, where a fine-tuned Qwen3-8B persona adapter stores attributes its context does not state. We ask how mu
Can reconstructed local evidence support complete claims about a model's expressed reasoning? We study the first and latest explicit scalar assignments to a que
Does selecting a precise answer from a known menu establish that an activation reader distinguishes every digit? We test prospective complete-amount prediction
Activation readers should preserve what a source expresses, including numerical errors, while assigning details to the requested reference. We study these requi
PersonaQA asks an Activation Oracle (AO) reader for an attribute that a fine-tuned Qwen3-8B stores about a persona, from one residual state. The task's readers
Activation differences between a fine-tuned and a base model are an established reader input, and a frozen Activation Oracle (AO) given such differences already
A base-model text reader of a 50-token read window describes Qwen3-8B reasoning traces with more trace-specific claims than Activation Oracle (AO) readers, but
Difference readers read a fine-tuned model's residual state minus a reference state on the same prompt; the reference is usually the base model, or a single qui
Can a frozen Activation Oracle (AO) recover chain-of-thought scalars from under $5\%$ of source positions without raw text, and what can pre-answer contextual a
Can an activation reader keep two variables apart when both queries receive exactly the same physical states? We test complete first/latest literal-value report
Does giving a frozen activation reader more source positions improve complete factual recovery, and does asking it for several records preserve that benefit? We
Can an activation reader reconstruct complete source records while also forecasting a model's actual next output, including mistakes? We study a controlled inve
Difference readers decode what a fine-tune was trained to do from its residual state minus a reference. We ask whether such readouts track the behaviour or the
Can a short training run make Activation Oracle readers forecast what a model writes next, beyond what the visible text gives? Earlier work found that two Qwen3
In paper 24's best list-free PersonaQA pipeline for Qwen3-8B (a trained Activation Oracle reader whose sampled answers are verified by the examined model's own
Can a model recover the complete set of facts it was taught without inventing missing fields? We study controlled fictional profiles, distinguishing injected as
When a language model emits a bare wrong binary decision ("No" or "Yes") on a multi-premise prompt without a reasoning trace, we study conditional operational r
Ordinary supervised adaptation can teach an Activation Oracle to report complete owner–time relations in a model's expressed computation. We train a three-layer