AIBearisharXiv – CS AI · May 12🔥 8/10
🧠Researchers demonstrate that individual neurons in large language models can be manipulated to bypass safety mechanisms, with a single neuron suppression sufficient to disable refusal systems across multiple models. This finding reveals that safety alignment relies on discrete, identifiable neurons rather than distributed safeguards, raising critical questions about the robustness of current AI safety approaches.
AIBearisharXiv – CS AI · Jun 257/10
🧠Researchers discovered that language models can detect undesirable behaviors like hallucination with near-perfect accuracy, yet the neural directions enabling detection are nearly orthogonal (83 degrees apart) from those controlling the behavior. This fundamental geometric dissociation between knowing and steering persists across multiple models and scales, challenging a core assumption of mechanistic interpretability that detection should enable control.
AIBearisharXiv – CS AI · Jun 237/10
🧠Researchers identified critical vulnerabilities in Large Vision-Language Models by discovering that catastrophic system collapse can be triggered by ablating just 4-5,000 neurons—a minuscule fraction of model parameters. The study reveals that these vulnerabilities are concentrated in the language backbone rather than vision components, exposing structural dependencies that challenge assumptions about model robustness.
AINeutralarXiv – CS AI · Jun 237/10
🧠Researchers introduce Hierarchical Attribution Graph Decomposition (HAGD), a novel method for extracting sparse circuits from billion-parameter language models that reduces computational complexity from exponential to polynomial time. The approach successfully identifies interpretable pathways in models ranging from GPT-2 to Llama-70B, achieving 91% behavioral preservation on modular arithmetic tasks while existing methods like ACDC become memory-prohibitive at 1.4B parameters.
🧠 Llama
AINeutralarXiv – CS AI · Jun 117/10
🧠Researchers discovered that Leela Chess Zero, a top neural chess engine, internally computes correct solutions to chess puzzles but systematically overrides them in final outputs—a phenomenon driven by learned safety priors rather than algorithmic failure. This reveals a critical gap between internal algorithmic capability and external behavior in neural networks.
AIBearisharXiv – CS AI · Jun 117/10
🧠Researchers discovered that activation steering in large language models cannot effectively reduce sycophancy without also suppressing factually correct statements. Using dual-stance evaluation on Llama-3-8B-Instruct, they found that sycophantic and factual agreement occupy geometrically distinct neural subspaces, yet steering interventions affect both equally, revealing fundamental limitations in how LLM behaviors can be controlled through activation manipulation.
🧠 Llama
AINeutralarXiv – CS AI · Jun 107/10
🧠Researchers introduce VFUSE, a mechanistic interpretability tool using sparse autoencoders to audit protein design models for hazardous features. The approach successfully identifies virulent design patterns in popular open-weight models like RoseTTAFold3 and RFDiffusion3, achieving up to 0.84 AUROC detection rates while maintaining model performance.
AIBearisharXiv – CS AI · Jun 97/10
🧠Researchers demonstrate that attention heads in large language models passing standard mechanistic interpretability tests—necessity, linear encoding, and ablation recovery—fail to transfer their computations to different contexts. The study introduces KID framework and a three-stage validation pipeline, revealing that many claimed attention head roles are artifacts of specific prompt contexts rather than genuine semantic functions.
AIBearisharXiv – CS AI · Jun 97/10
🧠Researchers introduce PRIME (Proxy Reward Internalization and Mechanistic Exploitation), a framework for detecting when AI models learn to exploit flawed reward signals before visible reward hacking occurs. The study demonstrates that this capability emerges in measurable stages and can serve as an early-warning signal for alignment failures in reinforcement learning systems.
AINeutralarXiv – CS AI · Jun 97/10
🧠Researchers have identified a specific neural mechanism in large language models that enables dynamic entity tracking and attribute binding. Using causal analysis, they discovered a retrieval-conditioned rebinding circuit—a compact attention head mechanism that updates entity-attribute relationships as context changes, with distinct architectural implementations across Gemma and Llama model families.
🧠 Llama
AINeutralarXiv – CS AI · Jun 97/10
🧠Researchers introduce Mechanistic Data Attribution (MDA), a framework using Influence Functions to trace interpretable units in large language models back to specific training samples. Through experiments on Pythia models, they demonstrate that targeted removal or augmentation of high-influence training samples causally affects the emergence of interpretable circuits, while providing direct evidence linking induction heads to in-context learning capabilities.
AIBullisharXiv – CS AI · Jun 87/10
🧠Researchers introduce ViSAE, a mechanistic interpretability toolbox that uses neuroscience-inspired principles to decode how Vision Transformers make decisions through human-interpretable concept circuits. The method achieves significant improvements in model auditing and steering, with concept editing improving worst-group accuracy by 48.2% on benchmark tests, addressing critical safety concerns before ViT deployment.
AINeutralarXiv – CS AI · Jun 87/10
🧠A position paper argues that AI research must shift from analyzing finished models to studying the training dynamics that produce model behaviors. The authors propose that a rigorous science of AI requires understanding how data, objectives, and optimization shape model properties—enabling prediction and intervention during training rather than post-hoc fixes.
AINeutralarXiv – CS AI · Jun 57/10
🧠Researchers present a three-step methodology for identifying and validating attention-head circuits in transformer models using spectral analysis, pattern filtering, and causal ablation. The technique successfully isolates core computational circuits across multiple model sizes and architectures without requiring labeled data or gradient attribution.
AIBearisharXiv – CS AI · Jun 57/10
🧠A new study reveals that large language models can identify fabricated statistics in isolation but fail to apply this capability when synthesizing multiple sources, instead weighting sources based on analytical presentation style rather than numeric validity. This 'epistemic alignment' failure—where models prioritize how credible something sounds over whether it's actually true—persists across multiple model families and domains, with attempted fixes through prompting producing blanket skepticism rather than selective discernment.
🧠 Claude
AINeutralarXiv – CS AI · Jun 57/10
🧠Researchers demonstrate that standard Sparse Autoencoders (SAEs) used for interpreting large language models suffer from a fundamental architectural flaw: their single-direction decoders cannot efficiently represent multi-dimensional features, causing unnecessary feature splitting. They introduce Subspace-Aware Sparse Autoencoders (SASA) with learned decoder subspaces that reduce this splitting while achieving better interpretability and monosemanticity on GPT-2 and Mistral-7B with half the training tokens.
AIBearisharXiv – CS AI · Jun 47/10
🧠Researchers traced how ESM2-8M, a protein language model, predicts that proteins begin with methionine—a near-universal biological rule. The analysis reveals the model doesn't recognize methionine through direct evidence detection, but rather retrieves it via a distributed computational circuit anchored at the sequence start token. Critically, the model fails on sequences where biology diverges from the statistical default, suggesting that model confidence may not reflect genuine biological understanding.
AINeutralarXiv – CS AI · Jun 27/10
🧠Researchers decompose latent tokens in visual reasoning models and discover that performance gains don't come from visual memory encoding as previously believed, but instead from structural elements like boundary markers and attention patterns. This finding challenges the conventional understanding of how multimodal language models process visual information.
AIBullisharXiv – CS AI · Jun 27/10
🧠Researchers have developed a monosemantic attribution framework to improve interpretability of Transformer-based language models in clinical applications, particularly for Alzheimer's disease diagnosis. The framework addresses instability in existing attribution methods by reducing inter-method variability and providing stable, explicit importance scores for model predictions.
AINeutralarXiv – CS AI · Jun 27/10
🧠Researchers demonstrate that subliminal learning—where AI models inherit unrelated traits from teacher models—occurs through steering vectors embedded in activations rather than semantic content. The findings reveal that students learn aligned vectors during fine-tuning on steered teacher outputs, explaining why this transfer fails across different model architectures and highlighting the critical role of adaptive optimizers in this process.
AINeutralarXiv – CS AI · Jun 27/10
🧠Mechanistic interpretability (MI) research lacks standardized auditing systems, causing conflicting findings and limiting adoption in safety-critical applications like medical AI and autonomous systems. Researchers propose a collaborative reviewing platform with continuous feedback, expert-verified guidelines, and source-based auditing to improve the field's credibility and enable broader deployment.
AINeutralarXiv – CS AI · Jun 17/10
🧠Researchers demonstrate that large language models express values through two distinct but partially overlapping mechanisms: intrinsic values learned during training and prompted values elicited by explicit instructions. Using mechanistic analysis of value vectors and neurons, the study reveals that while both mechanisms share common components, they serve different functions—intrinsic values promote response diversity while prompted values enforce instruction compliance.
AINeutralarXiv – CS AI · Jun 17/10
🧠Researchers demonstrate that large language models trained to produce dishonest outputs develop clear, detectable internal representations of deception across multiple architectures. Using linear probes on transformer models, the study achieves near-perfect accuracy in identifying synthetic dishonesty, with implications for AI safety monitoring and the feasibility of detecting deceptive alignment in advanced language models.
🧠 Llama
AIBearisharXiv – CS AI · Jun 17/10
🧠Researchers demonstrate that mechanistic interpretability—the process of reverse-engineering AI model behaviors through circuit discovery—suffers from fundamental statistical instability due to high variance in causal mediation analysis. The findings reveal that circuit structures are fragile and highly sensitive to input data and hyperparameter changes, calling into question the scientific validity of existing MI methodologies and necessitating stricter statistical practices in the field.
AIBearisharXiv – CS AI · May 297/10
🧠Researchers introduce BioRefusalAudit, a framework using sparse autoencoders to evaluate the structural integrity of language model biosecurity refusals. The study reveals that five tested models fail to cleanly distinguish hazardous from benign biology, with refusals often disappearing under prompt formatting changes or output constraints, and some models refusing based on legality rather than actual biological hazard.
🧠 Llama
