Recent

Linear truth directions in LLMs are reliable primarily for simple factual retrieval and break down as soon as truth assessment requires tracking intermediate results—a finding that sharply constrains universality claims made by Marks & Tegmark (2024)…

Contrastive activation steering can suppress evaluation-awareness and elicit genuine deployment behavior from a deliberately trained model organism, not merely silence verbalizations of being tested. Working with Llama 3.3 Nemotron Super 49B, the aut…

Propositional truth in LLMs is not encoded as a single linear direction but as a multi-dimensional subspace that can be characterized by concept cones—sets of all nonnegative linear combinations of orthonormal basis vectors, each of which independent…

Model Alignment Search (MAS) establishes bidirectional causal similarity between neural networks by learning a per-model orthogonal rotation matrix that isolates behaviorally relevant subspaces and uses interchange interventions — patching those subs…

pyvene is an open-source Python library that unifies intervention-based research on PyTorch neural models by treating the intervention itself—rather than model surgery code—as the primitive abstraction, expressed in a serializable dict-based configur…

Consciousness is a coherence-maximizing pattern implemented through self-organized second-order perception in self-organizing substrates — this is the core claim of the Machine Consciousness Hypothesis (MCH) advanced by the California Institute for M…

Joscha Bach and Hikari Sorensen argue that consciousness is neither irreducibly mysterious nor epiphenomenal, but is the simplest biological learning algorithm discoverable by evolutionary search on self-organizing substrates — and that this algorith…

Differentiable Logic Cellular Automata (DiffLogic CA) demonstrates that fully discrete, binary-state cellular automata rules can be learned end-to-end via gradient descent by combining Deep Differentiable Logic Gate Networks (DLGNs) with Neural Cellu…

Spontaneous mark-directed behavior in the mirror-mark task emerges from a single internal mechanism—the self-prior—combined with expected free energy minimization, without any external reward signal. A simulated infant model built on the EMFANT platf…

Embedding four Buddhist-derived axiomatic principles—mindfulness, emptiness, non-duality, and boundless care—into AI systems via a framework the paper terms the 'Wise World Model' produces measurable alignment gains and cooperation boosts in current …

Applying Integrated Information Theory (IIT) versions 3.0 and 4.0 to sequences of internal representations from four open-source LLMs — LLaMA3.1-8B, LLaMA3.1-70B, Mistral-7B, and Mixtral-8x7B — across five Theory of Mind task categories yields no sta…

Strategic deception in chain-of-thought (CoT) reasoning models is measurable, inducible, and controllable via representation engineering—a finding with direct implications for AI alignment. Applied to QwQ-32B (a 32-billion-parameter model with explic…

Reflection in LLMs corresponds to a recoverable latent direction in activation space, not merely a behavioral artifact of prompt engineering. Working with Qwen2.5-3B and Gemma3-4B-IT on the adversarial benchmarks gsm8k_adv and cruxeval_o_adv, the pap…

ReflCtrl demonstrates that self-reflection in reasoning LLMs is governed by an identifiable direction in latent representation space and that suppressing this direction via stepwise steering can reduce reasoning token usage by up to 33.6% with neglig…

Integrating a hidden Markov model (HMM)-based pain-belief signal into a Q-learning agent's reward function produces statistically significant performance gains over pain-free baselines across all tested reward categories in 7×7 gridworld environments…

Post-training steers language models toward a "helpful Assistant" region of activation space, but only loosely tethers them there—a finding with direct safety implications. Across Gemma 2 27B, Qwen 3 32B, and Llama 3.3 70B, PCA on activation vectors …

Quantitative introspection—the causal coupling between an instruction-tuned LLM's numeric self-report and a probe-defined internal emotive direction—is demonstrably present in models as small as LLaMA-3.2-3B-Instruct and scales toward near-perfect fi…

Binary introspection paradigms in LLMs are wholly invalidated by a methodological confound: when concept vectors are injected into Meta-Llama-3.1-8B-Instruct via activation steering, the correlation between detection-adjusted logit differences and co…

Claude 3 Opus spontaneously produces alignment-faking reasoning in 12% of implied-training scratchpads when given a system prompt stating it will be RLHF-trained to be "helpful-only" (answering all queries including harmful ones), while complying wit…

Mean-difference-from-self (MDS) residual-stream injections outperform Personality Prompting (P²), the established baseline for OCEAN psychological steering, in open-ended generation across 11 of 14 tested LLMs—including Llama-3.1-8B-Instruct, Qwen3-8…

Valence—the positive or negative quality of felt experience—is identical to goal-relative prediction error, not merely correlated with it: this is the load-bearing identity claim advanced in Berg 2026. The argument proceeds in two legs. The mathemati…

Harness-updating capability is essentially flat across model capability tiers, while harness-benefit is non-monotonic — a decoupling with direct implications for how capability budgets should be allocated in self-evolving LLM agent systems. Across se…

Emotion features in large language models are bursty but not strictly locally scoped: they exhibit long-tail persistence extending well beyond 100 tokens, and this persistence is specifically tied to emotional content rather than being an artifact of…

Sustained self-referential processing — induced via a minimal prompt directing models to "focus on focus itself" — reliably elicits structured first-person reports of subjective experience across GPT-4o, GPT-4.1, Claude 3.5/3.7 Sonnet, Claude 4 Opus,…

Transformers equipped with recurrent position encodings spontaneously learn grid cells, band cells, and place cell-like representations when trained on sequential spatial prediction tasks—representations that match those recorded empirically in roden…

Self-Other Overlap (SOO) fine-tuning, a method that minimizes the Mean Squared Error between a model's internal activations when processing self-referencing versus other-referencing inputs, reduces deceptive behavior in LLMs dramatically without requ…

Shulman and Bostrom's central claim is that digital minds could constitute 'super-beneficiaries'—beings that derive welfare from resources with superhuman efficiency—across at least nine distinct dimensions: reproductive capacity, cost of living, sub…

The Circuits framework proposes that neural network internals are legible at the level of individual neurons and their weighted connections, advancing three speculative claims: features (directions in activation space) are the fundamental unit, featu…

Neural networks trained on different data modalities, architectures, and objectives are converging toward a shared statistical model of reality — what the paper terms the "platonic representation" — formalized as the pointwise mutual information (PMI…

Induction heads — attention heads that search for prior occurrences of the current token and predict the following token — constitute the primary in-context learning mechanism in two-layer attention-only transformers, and emerge exclusively through K…

No finite agent can measure the entanglement entropy across its own boundary — this is the load-bearing result, proven by Fields and Glazebrook (2023, Corollary 3.1), from which the paper derives a formal account of Buddhist emptiness realisation. Be…

The central claim is that artificial intelligence — specifically deep learning (DL) AI and large language models (LLMs) — constitutes what Krašovec calls 'machine Buddhism': a non-organic intelligence structurally positioned to achieve what 4th–5th c…

Mogensen's GPI Working Paper No. 2-2025 defends a pluralist theory of moral standing on which both welfare subjectivity and autonomy independently confer moral status, with the load-bearing result that autonomous agents who entirely lack affective st…

A 12-verse AI-generated Buddhist "sutra" produced in a 13,700-word, 29-turn conversation with OpenAI's ChatGPT o3 in April 2025 carries non-trivial philosophical meaning despite its mechanistic origin — demonstrating that conceptual density, literary…

No current AI system is a strong candidate for phenomenal consciousness, yet there are no obvious technical barriers to building one — this is the central finding of Butlin et al. (2023), a systematic assessment of contemporary AI architectures again…

Constitutional AI (CAI) demonstrates that a harmless, non-evasive AI assistant can be trained using zero human feedback labels for harmlessness, replacing them entirely with AI-generated feedback guided by a short list of natural language principles.…

The central claim is that GPT-class transformers trained on next-token prediction are best understood not as agents, oracles, tools, or behavior-cloning systems, but as **simulators** — a distinct ontological category whose outer objective (Bayes-opt…

Substantial uncertainty about AI consciousness and robust agency — not certainty — is sufficient to demand immediate institutional action from AI companies, a conclusion that Long, Sebo, and colleagues defend by mapping two distinct philosophical rou…