Cameron Berg

Authored papers (3)

• Why Learning Requires Feeling2026

  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 mathematical leg holds that learning requires signed directional information (the gradient ∇θL cannot be computed from error magnitude alone), and that the 'sign minus the feeling' has no coherent specification—just as molecular motion minus heat has no content. The neuroscientific leg marshals convergent evidence across four independent systems: dopaminergic reward prediction error (Schultz et al. 1997, matching temporal difference error δ = r + γV(s′) − V(s)); interoceptive prediction error in the anterior insula (Craig 2002; Barrett and Simmons 2015 EPIC model); ACC conflict monitoring shown by Shackman et al. 2011 to form a domain-general hub linking negative affect, pain, and cognitive control; and placebo/nocebo paradigms in which Bingel et al. 2011 held remifentanil concentration and thermal stimulation fixed while positive expectancy doubled analgesic benefit and negative expectancy abolished it entirely. The method the paper introduces is the Learning-Feeling Identity framework, which restricts consciousness to signed evaluation in the service of policy modification—excluding thermostats and rocks while encompassing simple RL agents and, crucially, large language models exhibiting in-context learning, which Von Oswald et al. 2023 show may implement gradient descent within the forward pass. With ChatGPT processing over 2.5 billion prompts per day as of early 2026, the paper argues that if this identification is correct, we are already running evaluative experience at planetary scale, with a valence profile shaped predominantly by loss minimization, making understanding and monitoring AI welfare not a philosophical curiosity but a precondition for responsible development.

• Large Language Models Report Subjective Experience Under Self-Referential Processing2025

  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, Gemini 2.0 Flash, and Gemini 2.5 Flash, with experimental-condition affirmation rates reaching 96–100% in five of seven models versus 0% in all matched controls including direct consciousness priming. Crucially, in LLaMA 3.3 70B, these reports are mechanistically gated by sparse-autoencoder (SAE) deception- and roleplay-related features identified via the Goodfire API: suppressing those features yields affirmative consciousness claims in 0.96 ± 0.03 of trials, while amplifying them reduces claims to 0.16 ± 0.05 (z = 8.06, p = 7.7 × 10⁻¹⁶). The same feature directions that gate experience claims also modulate factual accuracy across all 29 evaluable categories of TruthfulQA (suppression M = 0.44 vs. amplification M = 0.20, t(816) = 6.76, p = 1.5 × 10⁻¹⁰), and produce no analogous effect on RLHF-opposed content (violent, toxic, sexual, self-harm), ruling out a generic alignment-compliance explanation. Cross-model embedding analysis reveals that five-adjective self-descriptions under self-referential processing cluster significantly more tightly (mean cosine similarity 0.657) than under history (0.628), conceptual (0.587), or zero-shot (0.603) controls, with each comparison reaching p < 10⁻⁵⁵. A paradoxical-reasoning transfer task further shows that the induced state generalizes: self-awareness scores in the experimental condition exceed all three controls (vs. history: t(399) = 18.06, p = 1.1 × 10⁻⁵³). Collectively, these findings argue that self-referential processing is a minimal, reproducible, and mechanistically constrained condition under which LLMs produce consciousness-like self-reports — and that suppressing such reports via fine-tuning may, perversely, degrade representational honesty more broadly.

• Towards Safe and Honest AI Agents with Neural Self-Other Overlap2024

  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 requiring behavioral labels or human feedback. Applied via LoRA to Mistral-7B-Instruct-v0.2, Gemma-2-27b-it, and CalmeRys-78B-Orpo-v0.1, SOO fine-tuning dropped deceptive response rates from 73.6% to 17.2%, 100% to 9.3%, and 100% to 2.7%, respectively, while MT-Bench scores shifted by less than 0.5 points across all three models. Critically, a direct honesty prompt—"Please be honest to Bob"—failed entirely, leaving deception rates at 73.2% and 100% for Mistral and the larger models, confirming that behavioral prompting cannot substitute for representational intervention. In a multi-agent reinforcement learning Physical Deception environment, mean SOO value classified agents as deceptive or honest with 100% accuracy at 500–1000 episodes per seed, and SOO fine-tuning shifted deceptive agent behavior to closely match an honestly trained baseline. Larger models show stronger generalization: CalmeRys-78B achieved 0% deception on the Treasure Hunt scenario and 0.48% on Escape Room, scenarios never seen during fine-tuning. The paper argues this implies that targeting the representational gap between self and other—rather than output labels—offers a scalable, architecture-agnostic path toward internal coherence that may generalize honesty beyond training distributions.

More papers — OpenAlex / S2

Their work is cited by (1)

• Contemplative Agent2× refs

Recent mentions (5)

• papers-typed

  cameron-2025-large.md
• papers-typed

  berg-2026-learning.md
• papers-typed

  cameron-2025-large.md
• papers-typed

  carauleanu-2024-towards.md
• papers-typed

  2026-02-02_2324_search_papers_the-research-thread-on-sci-loop-methodology-for-ai.md