Dr. No steers sparse autoencoder features at inference to make vision-language and language models abstain when evidence is insufficient, beating post-hoc rejection on several abstention benchmarks, with gains source-dependent.
Probes, an SAE and optimized directions in small LMs' hidden states predict refusal versus confabulation; pushing along them flips most confabulations into refusal openings.
The paper shows that a VLM's appropriate location-disclosure granularity can be read from hidden states, and proposes an inference-time router and steering method that improves privacy-aligned disclosure while keeping general capabilities largely intact.
Probes and steers 25 open-weight LLMs scoring essays, finding rubric directions share a dominant lever, so scores aren't independently controllable by rubric despite human agreement.
Adds a symmetry-based Gini regularizer to sparse autoencoder training to reduce representation blind spots in safety-critical text, improving weak-tail perturbation responses across GPT-2, Gemma 2 and Qwen 2.5.
Uses mechanistic interpretability to locate an attention head in medical chest X-ray ViTs that attends to annotation text, a spurious shortcut that output-level evaluation fails to detect.
Proposes a deception detector reading all layers of a language model's activations, with calibrated thresholds and ensemble-based uncertainty, evaluated on unseen deceptive tasks across Qwen3 model sizes.
The paper finds causally used clinical concept representations in eleven open-weight LLMs, shows steering them improves clinical performance, and tests robustness under adversarial role priming.
Proposes DiDAE, a fast diffusion-based method for disentangled counterfactual explanations of frozen vision foundation models, used to find and repair spurious classifier strategies.
Mechanistic analysis shows LLMs generate multi-valued relations in a canonical order, so prompts that deviate from it yield markedly less complete sets.
Uses causal mechanistic analysis to find a sparse circuit of attention heads and MLPs behind LLM overcommitment to answers, and trains a lightweight policy on it to improve abstention decisions.
Proposes linear probes on frozen embedder query representations to score medical, psychological, and topic risk in patient queries as guardrails, finding them competitive with LLM graders at lower latency.
Proposes a five-step diagnostic protocol for sparse-neuron localization claims and tests it on H-neurons for LLM hallucination, finding detection replicates but neuron selection is not uniquely localized.
Compares probing, attribution patching, SAEs and ablation to trace where stereotype information sits in Llama, Qwen and Gemma, finding probe and attribution peaks separate and language-agnostic features are rare.
Shows a refusal direction found in transformers transfers via rotation alignment to state-space and other architectures, where it is read at each layer's write site and steered to lower jailbreak success.
Analyzes how typos, shuffling and HotFlip perturbations propagate through GPT-2 and Qwen2.5 across behavior, hidden-state geometry and attention heads, finding single metrics can mislead robustness claims.
Patching experiments on three vision-language models find that mid-layer encoded answers are causally inactive in POPE hallucination errors, and errors split into three learnable failure modes.
Finds LLMs internally represent structural impossibility via a linear direction nearly orthogonal to the safety-refusal direction, so confident answers to unanswerable questions reflect a routing failure. Steering along it changes abstention behaviour.
Uses sparse autoencoders on pathology foundation model embeddings to find and mask domain-specific features, improving cross-domain robustness for prostate cancer relapse prediction.
Proposes a mechanistic explanation for why adversarial training reduces superposition in neural networks, tracing it to abandoning non-robust features and thus fewer features to represent.
Probes how demographic identity is represented, read out, and causally used in Mistral-7B for survey simulation, finding that faithful internal geometry does not translate into causal use.
Analyses LLM hidden-state geometry as small-world graphs and uses the resulting topology to flag hallucinations in retrieval-augmented generation without supervision.
Uses probes and activation patching to show that stylistic cues like emojis and grammar shift an LLM's internal estimates of user demographics, which changes its responses, raising privacy and manipulation concerns.
Probing and patching 18 open-source LLMs shows mythological knowledge is represented internally but lost at readout, and the failure depends on prompt language.
Shows natural-language activation explanations can pass reconstruction-based faithfulness tests while hiding false claims, and proposes RECAP, co-trained probe heads that keep designated content decodable and checkable against a probe.
Shows that linguistic rewrites shift LLM political stance judgments, and uses activation patching to localize the restoring signal to mid-to-late decoder layers at the final prompt position.
A survey of empirical AI consciousness research, covering interpretability studies of self-reports, emotion and valence vectors, and a pain axis that disrupts harm avoidance when steered.
Proposes a competition to use models' internal mechanisms and adversarial robustness tests to tell robust reasoning from spurious shortcuts in frontier math models, creating an open benchmark.
Probes on final-token activations estimate entity familiarity in Polish and other LLMs, and adding a familiarity direction steers refusal rates in Gemma-4-12B in both directions.
Uses mechanistic interpretability to show LLM-as-judge biases form type-specific activation subspaces that can be steered to shift scores and predict judge failures on unseen benchmarks.
Mechanistic analysis finds a middle-layer visual relay window in VLMs, and TRACE, an inference-time control method, improves evidence-grounded reasoning across benchmarks, though not framed around safety.
Argues that LLM functional welfare interventions must coherently shift welfare channels while preserving goal-monitoring, surveys options, and proposes a synthetic document fine-tuning experiment to reverse pathological failure responses for safety and moral reasons.
Shows that vision-language models defer to news outlet identity over article content, localizes this credibility prior mechanistically with sparse autoencoders, and reduces the override by steering.
Maps LLM internal features into functional parcels and links them to failures like hallucination, bias, refusal failure and sycophancy, enabling mechanism-guided detection and intervention, and to human cortical responses.
Proposes steering vectors to control LLM behavior and latent-space model calibrators to judge when outputs can be trusted, using model internals.
Linear probing and causal interventions on driving policies show they often fail to form timely predictions of nearby vehicles during near-collisions, and correcting those predictions yields safer plans.
Shows that LLM verbal confidence predicts whether a model commits to an answer better than whether the answer is correct, and that its internal signal can be steered to shift abstention.
Shows that framing changes LLM responses in mental-health prompts, that probes decode framing from internal representations, and that steering along those directions partly shifts behaviour.
Uses sparse autoencoders to show that OOD inputs, including typos and jailbreak prompts, activate more fallacious concepts in LLM internals, and proposes a mechanistically guided fine-tuning method for robustness.
Uses persistent homology on hidden-state point clouds to classify LLM ill-posedness and steers activations toward abstention or clarification, improving steering performance on two datasets across four model families.