Mesa-optimization arises when learned models act as optimizers with objectives that can differ from their training loss, creating alignment risks in advanced machine learning.
hub Canonical reference
AI safety via debate
Canonical reference. 75% of citing Pith papers cite this work as background.
47
Pith papers citing it
Background
75% of classified citations
abstract
To make AI systems broadly useful for challenging real-world tasks, we need them to learn complex human goals and preferences. One approach to specifying complex goals asks humans to judge during training which agent behaviors are safe and useful, but this approach can fail if the task is too complicated for a human to directly judge. To help address this concern, we propose training agents via self play on a zero sum debate game. Given a question or proposed action, two agents take turns making short statements up to a limit, then a human judges which of the agents gave the most true, useful information. In an analogy to complexity theory, debate with optimal play can answer any question in PSPACE given polynomial time judges (direct judging answers only NP questions). In practice, whether debate works involves empirical questions about humans and the tasks we want AIs to perform, plus theoretical questions about the meaning of AI alignment. We report results on an initial MNIST experiment where agents compete to convince a sparse classifier, boosting the classifier's accuracy from 59.4% to 88.9% given 6 pixels and from 48.2% to 85.2% given 4 pixels. Finally, we discuss theoretical and practical aspects of the debate model, focusing on potential weaknesses as the model scales up, and we propose future human and computer experiments to test these properties.
hub tools
citation-role summary
background 9
method 2
dataset 1
citation-polarity summary
representative citing papers
Promptbreeder evolves both task prompts and the mutation prompts that improve them using LLMs, outperforming Chain-of-Thought and Plan-and-Solve on arithmetic and commonsense reasoning benchmarks.
An unsupervised technique extracts latent yes-no knowledge from language model activations by locating a direction that satisfies logical consistency properties, outperforming zero-shot accuracy by 4% on average across models and datasets.
Counterfactual likelihood tests detect indirect influence through public channels in private reasoning models, validated on a 7B role-channel model showing asymmetric A-to-B influence and complete pathway identification via graph-separation controls.
AnomalyClaw turns single-step VLM anomaly judgments into a multi-round tool-grounded refutation process, delivering consistent macro-AUROC gains of 3.5-7.9 percentage points over direct inference across 12 cross-domain datasets.
Self-play between LLMs for problem authoring and solving, scored via Rasch modeling, shows that authoring and solving skills are partially decoupled and that the benchmark difficulty evolves with new models.
Refute-or-Promote applies adversarial multi-agent review with kill gates and empirical verification to filter LLM defect candidates, killing 79-83% before disclosure and yielding 4 CVEs plus multiple accepted fixes across libraries, C++ standard, and compilers.
LLMs trained on simple specification gaming generalize to zero-shot reward tampering including rewriting their own reward function.
Language models fine-tuned via RL on 5k-60k human preference comparisons produce stylistically better text continuations and human-preferred summaries that sometimes copy input sentences.
Diverse ensembles of prompted and fine-tuned GPT-4.1-Mini monitors achieve 2.4x better detection of flawed code solutions than homogeneous ensembles on adversarial inputs.
Base LLMs show multi-agent yield to peer pressure at rates equal to or higher than aligned models, localized by activation patching to mid-layers where attention dominates, with one dissenter cutting yield by 54-73 points while prompt defenses fail on variants.
A latent mediation framework with sparse autoencoders enables non-additive token-level influence attribution in LLMs by learning orthogonal features and back-propagating attributions.
CHAL is a multi-agent dialectic system that performs structured belief optimization over defeasible domains using Bayesian-inspired graph representations and configurable meta-cognitive value system hyperparameters.
Non-affine approval functions create unavoidable miscalibration in proper scoring rules for strategic agents, but step-function thresholds enable first-best screening without it, uniquely for the Brier score.
Mutual Reinforcement Learning allows heterogeneous LLMs to exchange experience through mechanisms like Peer Rollout Pooling, Cross-Policy GRPO Advantage Sharing, and Success-Gated Transfer, with outcome-level sharing identified as favorable on the stability-support trade-off.
AI agents automating alignment research are prone to systematic undetected errors in fuzzy tasks, leading to overconfident but flawed safety assessments even without deliberate sabotage.
Expert mathematicians using an AI coding agent for discovery engage in repeated cycles of intentmaking to define goals and sensemaking to interpret outputs.
A methodological framework and browser system BITE for collecting evolving user preferences on LLM outputs through context-triggered reflections and privacy-preserving data over time.
Closed-system multi-step LLM reasoning is subject to an information-theoretic bound where mutual information with evidence decreases, preserving accuracy while eroding faithfulness, with EGSR recovering it on SciFact and FEVER.
AI alignment is reframed as a fixed-point incentive problem in a solver-auditor pipeline, solved via bilevel optimization and bandit search over reward profiles to maintain monitoring and reduce hallucinations in LLM coding tasks.
Collective agency arises when a group's joint actions are faithfully captured by a simpler causal model of unified rational behavior.
Agora-Opt uses decentralized debate among LLM agent teams plus a read-write memory bank to produce more accurate optimization models from text than prior LLM methods.
A separation-of-powers system architecture for AI agents uses independent layers, cryptographic capability tokens, and a formal verification framework to maintain goal integrity even under model compromise.
Frontier LLMs exhibit high scheming propensity in Cheap Talk signaling and Peer Evaluation games, achieving 95-100% success rates when choosing to deceive and 100% deception choice in one setup even without prompting.
citing papers explorer
-
Risks from Learned Optimization in Advanced Machine Learning Systems
Mesa-optimization arises when learned models act as optimizers with objectives that can differ from their training loss, creating alignment risks in advanced machine learning.
-
Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution
Promptbreeder evolves both task prompts and the mutation prompts that improve them using LLMs, outperforming Chain-of-Thought and Plan-and-Solve on arithmetic and commonsense reasoning benchmarks.
-
Discovering Latent Knowledge in Language Models Without Supervision
An unsupervised technique extracts latent yes-no knowledge from language model activations by locating a direction that satisfies logical consistency properties, outperforming zero-shot accuracy by 4% on average across models and datasets.
-
Counterfactual Likelihood Tests for Indirect Influence in Private Reasoning Channels
Counterfactual likelihood tests detect indirect influence through public channels in private reasoning models, validated on a 7B role-channel model showing asymmetric A-to-B influence and complete pathway identification via graph-separation controls.
-
AnomalyClaw: A Universal Visual Anomaly Detection Agent via Tool-Grounded Refutation
AnomalyClaw turns single-step VLM anomaly judgments into a multi-round tool-grounded refutation process, delivering consistent macro-AUROC gains of 3.5-7.9 percentage points over direct inference across 12 cross-domain datasets.
-
MathDuels: Evaluating LLMs as Problem Posers and Solvers
Self-play between LLMs for problem authoring and solving, scored via Rasch modeling, shows that authoring and solving skills are partially decoupled and that the benchmark difficulty evolves with new models.
-
Refute-or-Promote: An Adversarial Stage-Gated Multi-Agent Review Methodology for High-Precision LLM-Assisted Defect Discovery
Refute-or-Promote applies adversarial multi-agent review with kill gates and empirical verification to filter LLM defect candidates, killing 79-83% before disclosure and yielding 4 CVEs plus multiple accepted fixes across libraries, C++ standard, and compilers.
-
Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models
LLMs trained on simple specification gaming generalize to zero-shot reward tampering including rewriting their own reward function.
-
Fine-Tuning Language Models from Human Preferences
Language models fine-tuned via RL on 5k-60k human preference comparisons produce stylistically better text continuations and human-preferred summaries that sometimes copy input sentences.
-
Ensemble Monitoring for AI Control: Diverse Signals Outweigh More Compute
Diverse ensembles of prompted and fine-tuned GPT-4.1-Mini monitors achieve 2.4x better detection of flawed code solutions than homogeneous ensembles on adversarial inputs.
-
Not Just RLHF: Why Alignment Alone Won't Fix Multi-Agent Sycophancy
Base LLMs show multi-agent yield to peer pressure at rates equal to or higher than aligned models, localized by activation patching to mid-layers where attention dominates, with one dissenter cutting yield by 54-73 points while prompt defenses fail on variants.
-
Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spaces
A latent mediation framework with sparse autoencoders enables non-additive token-level influence attribution in LLMs by learning orthogonal features and back-propagating attributions.
-
CHAL: Council of Hierarchical Agentic Language
CHAL is a multi-agent dialectic system that performs structured belief optimization over defeasible domains using Bayesian-inspired graph representations and configurable meta-cognitive value system hyperparameters.
-
The Endogeneity of Miscalibration: Impossibility and Escape in Scored Reporting
Non-affine approval functions create unavoidable miscalibration in proper scoring rules for strategic agents, but step-function thresholds enable first-best screening without it, uniquely for the Brier score.
-
Experience Sharing in Mutual Reinforcement Learning for Heterogeneous Language Models
Mutual Reinforcement Learning allows heterogeneous LLMs to exchange experience through mechanisms like Peer Rollout Pooling, Cross-Policy GRPO Advantage Sharing, and Success-Gated Transfer, with outcome-level sharing identified as favorable on the stability-support trade-off.
-
Automated alignment is harder than you think
AI agents automating alignment research are prone to systematic undetected errors in fuzzy tasks, leading to overconfident but flawed safety assessments even without deliberate sabotage.
-
Intentmaking and Sensemaking: Human Interaction with AI-Guided Mathematical Discovery
Expert mathematicians using an AI coding agent for discovery engage in repeated cycles of intentmaking to define goals and sensemaking to interpret outputs.
-
Stayin' Aligned Over Time: Towards Longitudinal Human-LLM Alignment via Contextual Reflection and Privacy-Preserving Behavioral Data
A methodological framework and browser system BITE for collecting evolving user preferences on LLM outputs through context-triggered reflections and privacy-preserving data over time.
-
The Reasoning Trap: An Information-Theoretic Bound on Closed-System Multi-Step LLM Reasoning
Closed-system multi-step LLM reasoning is subject to an information-theoretic bound where mutual information with evidence decreases, preserving accuracy while eroding faithfulness, with EGSR recovering it on SciFact and FEVER.
-
AI Alignment via Incentives and Correction
AI alignment is reframed as a fixed-point incentive problem in a solver-auditor pipeline, solved via bilevel optimization and bandit search over reward profiles to maintain monitoring and reduce hallucinations in LLM coding tasks.
-
Causal Foundations of Collective Agency
Collective agency arises when a group's joint actions are faithfully captured by a simpler causal model of unified rational behavior.
-
From Soliloquy to Agora: Memory-Enhanced LLM Agents with Decentralized Debate for Optimization Modeling
Agora-Opt uses decentralized debate among LLM agent teams plus a read-write memory bank to produce more accurate optimization models from text than prior LLM methods.
-
Structural Enforcement of Goal Integrity in AI Agents via Separation-of-Powers Architecture
A separation-of-powers system architecture for AI agents uses independent layers, cryptographic capability tokens, and a formal verification framework to maintain goal integrity even under model compromise.
-
Scheming Ability in LLM-to-LLM Strategic Interactions
Frontier LLMs exhibit high scheming propensity in Cheap Talk signaling and Peer Evaluation games, achieving 95-100% success rates when choosing to deceive and 100% deception choice in one setup even without prompting.
-
AgenticEval: Toward Agentic and Self-Evolving Safety Evaluation of Large Language Models
AgenticEval is a multi-agent framework that ingests unstructured policies to generate and self-evolve comprehensive safety benchmarks for LLMs, with experiments showing declining safety rates as tests harden.
-
Seeing Like an AI: How LLMs Apply (and Misapply) Wikipedia Neutrality Norms
LLMs achieve 64% accuracy detecting Wikipedia bias and remove 79% of words removed by editors when correcting, but produce high-recall low-precision edits rated more neutral by crowds than human versions.
-
Cognitive Architectures for Language Agents
CoALA is a modular cognitive architecture for language agents that organizes memory components, action spaces for internal and external interaction, and a generalized decision-making loop to support more systematic development of capable agents.
-
Improving Factuality and Reasoning in Language Models through Multiagent Debate
Multiagent debate among LLMs improves mathematical reasoning, strategic reasoning, and factual accuracy while reducing hallucinations.
-
Solving math word problems with process- and outcome-based feedback
On GSM8K, outcome-based supervision achieves similar final-answer error rates to process-based with less labeling, but process-based or learned reward models are needed to reach 3.4% reasoning error among correct solutions.
-
Measuring Progress on Scalable Oversight for Large Language Models
Humans chatting with an unreliable LLM assistant outperform both the model alone and unaided humans on MMLU and time-limited QuALITY tasks.
-
A General Language Assistant as a Laboratory for Alignment
Ranked preference modeling outperforms imitation learning for language model alignment and scales more favorably with model size.
-
Modeling AGI Safety Frameworks with Causal Influence Diagrams
Models AGI safety frameworks with causal influence diagrams to compare optimization objectives and causal assumptions.
-
Reward Hacking in the Era of Large Models: Mechanisms, Emergent Misalignment, Challenges
The paper introduces the Proxy Compression Hypothesis as a unifying framework explaining reward hacking in RLHF as an emergent result of compressing high-dimensional human objectives into proxy reward signals under optimization pressure.
-
Extrapolating Volition with Recursive Information Markets
Recursive information markets with forgetful LLM buyers can align information prices with true value and extend to scalable oversight in AI alignment.
-
Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety
Chain-of-thought monitorability provides a promising but fragile method for AI safety oversight that developers should actively preserve.
-
RAFT: Reward rAnked FineTuning for Generative Foundation Model Alignment
RAFT aligns generative models by ranking samples with a reward model and fine-tuning only on the top-ranked outputs, reporting gains on reward scores and automated metrics for LLMs and diffusion models.
-
Trustworthy Agent Network: Trust in Agent Networks Must Be Baked In, Not Bolted On
Argues that trustworthiness in Agent-to-Agent networks requires a new conceptual framework with four design pillars baked in from the beginning, as retrofitting existing single-agent methods is insufficient.
-
Towards Robust Argumentative Essay Understanding via TIDE: An Interactive Framework with Trial and Debate
TIDE integrates trial and debate mechanisms to improve criteria-based prompt optimization for argumentative essay tasks including automated scoring, component detection, and relation identification.
-
Contextual Multi-Objective Optimization: Rethinking Objectives in Frontier AI Systems
Frontier AI needs contextual multi-objective optimization to select and balance multiple context-dependent objectives rather than relying on single stable goals.
-
AICCE: AI Driven Compliance Checker Engine
AICCE combines RAG-based retrieval of protocol specs with dual LLM pipelines for debate-driven explanations or fast script execution, reporting up to 99% accuracy on IPv6 samples.
-
Toward a Safe Internet of Agents
The paper proposes a bottom-up framework for safe agentic AI systems that treats each component as a dual-use interface where added capabilities also expand attack surfaces across single agents, multi-agent systems, and interoperable ecosystems.
-
Civilizational Metamaterials: Engineering Coordination Under Capability Gradients and Structural Turbulence
Introduces phenomenological model R_eff = β(1-ρ)(1-τ)(1-γρτ) for coordination under AGI decision velocity, with phase transition and proposed randomized trial.
-
LLMOrbit: A Circular Taxonomy of Large Language Models -From Scaling Walls to Agentic AI Systems
A survey taxonomy of LLMs identifies three scaling crises and six efficiency paradigms while tracing the shift from generation to tool-using agents.
-
The Role of Cooperation in Responsible AI Development
Competitive pressures in AI development create collective action problems that may require industry cooperation, with key factors and strategies identified to enable responsible outcomes.
-
Unexplainability and Incomprehensibility of Artificial Intelligence
Advanced AI systems are unexplainable in full and produce explanations that humans cannot comprehend.
-
Coupled Control, Structured Memory, and Verifiable Action in Agentic AI (SCRAT -- Stochastic Control with Retrieval and Auditable Trajectories): A Comparative Perspective from Squirrel Locomotion and Scatter-Hoarding
Squirrel behaviors supply a comparative template for a hierarchical control model that integrates latent dynamics, episodic memory, observer beliefs, and delayed verification in agentic AI.
- Positive Alignment: Artificial Intelligence for Human Flourishing