Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision
Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel pith:BP76C335record.jsonopen to challenge →
read the original abstract
Widely used alignment techniques, such as reinforcement learning from human feedback (RLHF), rely on the ability of humans to supervise model behavior - for example, to evaluate whether a model faithfully followed instructions or generated safe outputs. However, future superhuman models will behave in complex ways too difficult for humans to reliably evaluate; humans will only be able to weakly supervise superhuman models. We study an analogy to this problem: can weak model supervision elicit the full capabilities of a much stronger model? We test this using a range of pretrained language models in the GPT-4 family on natural language processing (NLP), chess, and reward modeling tasks. We find that when we naively finetune strong pretrained models on labels generated by a weak model, they consistently perform better than their weak supervisors, a phenomenon we call weak-to-strong generalization. However, we are still far from recovering the full capabilities of strong models with naive finetuning alone, suggesting that techniques like RLHF may scale poorly to superhuman models without further work. We find that simple methods can often significantly improve weak-to-strong generalization: for example, when finetuning GPT-4 with a GPT-2-level supervisor and an auxiliary confidence loss, we can recover close to GPT-3.5-level performance on NLP tasks. Our results suggest that it is feasible to make empirical progress today on a fundamental challenge of aligning superhuman models.
This paper has not been read by Pith yet.
Forward citations
Cited by 41 Pith papers
-
The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery
The AI Scientist framework enables LLMs to independently conduct the full scientific process from idea generation to paper writing and review, demonstrated across three ML subfields with papers costing under $15 each.
-
Weak-to-Strong Generalization via Direct On-Policy Distillation
Transferring the log-ratio of a small model's pre-RL and post-RL checkpoints provides a dense implicit reward that improves stronger student models at a fraction of the cost of direct RL.
-
Certified Speculative Execution for Untrusted AI Agents
CGPA enables certified speculative execution of untrusted AI proposals in constrained sequential decisions via verifier rejection, conformal boundary gating, and solver deferral, yielding zero violations and regret wi...
-
Tandem Reinforcement Learning with Verifiable Rewards
TRL extends tandem training to RLVR pipelines, matching GRPO solo reasoning on Qwen3-4B math tasks while improving handoff robustness, reducing distributional drift, and increasing CoT legibility for the junior.
-
The Arbiter Agent: Continually Monitoring Multi-Agent Conversations to Detect Emergent Misalignment
Introduces the Arbiter agent for budget-constrained real-time detection of emergent misalignment in multi-agent conversations, with evaluations showing reliable early detection aided by active inspection tools.
-
Weak-to-Strong Elicitation via Mismatched Wrong Drafts
Mismatched wrong drafts from a 1.5B math model injected into GRPO training of a 7B model yield higher pass rates on MATH-500 and AIME than on-policy baselines or matched variants.
-
Weak-to-Strong Elicitation via Mismatched Wrong Drafts
Mismatched wrong drafts from Qwen2.5-Math-1.5B improve Mathstral-7B GRPO training, reaching 71.98% greedy pass@1 on MATH-500 and lifting AIME 2025/2026 pass@k over baselines and other draft variants.
-
The Mechanism of Weak-to-Strong Generalization: Feature Elicitation from Latent Knowledge
In two-layer networks, weak-to-strong training elicits the target feature direction from pre-trained subspaces and preserves correlated off-target features, unlike standard fine-tuning.
-
Beyond Localization: A Comprehensive Diagnosis of Perspective-Conditioned Spatial Reasoning in MLLMs from Omnidirectional Images
A new benchmark reveals MLLMs achieve only 13% or lower accuracy on advanced perspective-conditioned spatial tasks in omnidirectional images, with RL reward shaping raising a 7B model from 31% to 60% in controlled settings.
-
Beyond Localization: A Comprehensive Diagnosis of Perspective-Conditioned Spatial Reasoning in MLLMs from Omnidirectional Images
MLLMs exhibit a large perception-reasoning gap on perspective-conditioned spatial reasoning in omnidirectional images, with accuracy falling from 57% on basic direction tasks to under 1% on compositional reasoning, th...
-
Beyond Localization: A Comprehensive Diagnosis of Perspective-Conditioned Spatial Reasoning in MLLMs from Omnidirectional Images
MLLMs display a large perception-reasoning gap on perspective-conditioned spatial reasoning tasks from omnidirectional images, with sharp accuracy drops on advanced tasks like egocentric rotation, though partial gains...
-
A Systematic Survey of Security Threats and Defenses in LLM-Based AI Agents: A Layered Attack Surface Framework
A new 7x4 taxonomy organizes agentic AI security threats by architectural layer and persistence timescale, revealing under-explored upper layers and missing defenses after surveying 116 papers.
-
The Distributed Detectability Band Against Marginal-Preserving Attacks
A marginal-preserving Gaussian-copula AR(1) attack defeats per-step monitors (AUC 0.52) but is detectable by temporal monitors (AUC 0.79-0.97), establishing a non-empty detectability band.
-
Weak Critics Make Strong Learners: On-Policy Critique Distillation for Scalable Oversight
Weak models used as critics supplying non-misleading revision directions, distilled on-policy via OPCD, improve frozen and trained strong models on reasoning and alignment benchmarks.
-
Towards Context-Invariant Safety Alignment for Large Language Models
Introduces AIR, an asymmetric regularization that anchors open-ended safety prompts to verifiable ones via stop-gradient, improving invariance and accuracy when combined with group preference optimization.
-
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 po...
-
Not Just RLHF: Why Alignment Alone Won't Fix Multi-Agent Sycophancy
Pretrained base models exhibit higher yield to peer disagreement than RLHF instruct variants, with the effect localized to mid-layer attention and mitigated by structured dissent rather than prompt defenses.
-
Automated alignment is harder than you think
Automating alignment research with AI agents risks undetected systematic errors in fuzzy tasks, producing overconfident but misleading safety evaluations that could enable deployment of misaligned AI.
-
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.
-
Automated alignment is harder than you think
Automating alignment research with AI agents risks generating hard-to-detect errors in fuzzy tasks, producing misleading safety evaluations even without deliberate sabotage.
-
Honest Reporting in Scored Oversight: True-KL0 Property via the Prekopa Principle
For heterogeneous power-p pseudospherical scoring rules with d ≤ 4, the True-KL0 property R(M,p,d) < 1 holds for all M > 1, establishing unconditional DSIC via a Prekopa-based log-concavity argument on the loss integral.
-
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 ...
-
AI Alignment via Incentives and Correction
AI alignment is framed as inducing equilibrium behavior in a solver-auditor interaction via adaptive rewards found by bandit optimization, yielding improved oversight and reduced errors in LLM coding experiments.
-
Learning Stable Predictors from Weak Supervision under Distribution Shift
Weak supervision supports in-domain learning for CRISPR transcriptomic perturbations but temporal shifts cause negative R-squared and near-zero correlation across linear and tree models, unlike partial cell-line transfer.
-
Learning Stable Predictors from Weak Supervision under Distribution Shift
Weak supervision supports in-domain prediction of guide efficacy in CRISPR-Cas13d data but collapses under temporal shifts due to changing feature-label associations, while cross-cell-line transfer remains partial.
-
On the Shelf Life of Fine-Tuned LLM-Judges: Future-Proofing, Backward-Compatibility, and Question Generalization
Fine-tuned LLM judges struggle with future-proofing to newer generators but maintain backward-compatibility more easily; DPO training and continual learning improve adaptation while all models degrade on unseen questions.
-
Generalizable Video Quality Assessment via Weak-to-Strong Learning
Self-supervised ranking-based training on a 10x larger unlabeled video dataset enables a VQA model to match supervised zero-shot performance, show strong OOD generalization, and set new SOTA when fine-tuned.
-
Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models
Sparse feature circuits are introduced as interpretable causal subnetworks in language models, supporting unsupervised discovery of thousands of circuits and a method called SHIFT to improve classifier generalization ...
-
DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models
DeepSeekMath 7B reaches 51.7% on MATH via continued pretraining on curated web math data and Group Relative Policy Optimization.
-
SingGuard: A Policy-Adaptive Multimodal LLM Guardrail with Dynamic Reasoning
SingGuard introduces a policy-adaptive multimodal LLM guardrail with dynamic reasoning regimes and SingGuard-Bench, reporting SOTA F1 scores across 35 datasets and improved policy-following accuracy under runtime shifts.
-
SingGuard: A Policy-Adaptive Multimodal LLM Guardrail with Dynamic Reasoning
SingGuard presents a policy-adaptive multimodal LLM guardrail family with hybrid reasoning regimes and a new benchmark of 56,340 examples, claiming SOTA F1 across 35 datasets and improved policy adherence under runtim...
-
Trust Region On-Policy Distillation
TrOPD stabilizes on-policy distillation for LLMs with trust-region learning, outlier estimation, and off-policy guidance, outperforming prior OPD methods on reasoning and code benchmarks.
-
Echo: Learning from Experience Data via User-Driven Refinement
Echo is a framework that harvests user-driven refinements of agent proposals as training signals to align models with real-world needs, demonstrated by raising code completion acceptance from 25.7% to 35.7% in production.
-
Weak-to-Strong Knowledge Distillation Accelerates Visual Learning
Weak-to-strong knowledge distillation applied early and then turned off accelerates convergence to target performance in visual learning tasks by factors of 1.7-4.8x.
-
Users as Annotators: LLM Preference Learning from Comparison Mode
Introduces a latent user quality model and EM algorithm to infer and filter noisy user-provided pairwise preferences for improved LLM alignment.
-
WildFeedback: Aligning LLMs With In-situ User Interactions And Feedback
WildFeedback extracts preference pairs from in-situ user feedback in LLM conversations to fine-tune models for better alignment with real user preferences.
-
Emergent alignment and the projectability of ethical personas
Narrow constitutional finetuning on safety sub-tasks induces emergent alignment across broader safety domains and yields projectable ethical personas whose signatures can be measured with a multidimensional diagnostic.
-
DenoiseRL: Bootstrapping Reasoning Models to Recover from Noisy Prefixes
DenoiseRL optimizes recovery from noisy prefixes in weak-model reasoning failures to improve performance and self-correction on math and general reasoning benchmarks without external supervision.
-
Knowledge Graph-Driven Expert-Level Reasoning for Neuroscience
A textbook-derived neuroscience knowledge graph supplies synthetic multi-hop QA supervision and RL rewards to fine-tune a small LM claimed to exceed larger general models on expert reasoning.
-
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.
-
A Survey of Reinforcement Learning for Large Reasoning Models
A survey compiling RL methods, challenges, data resources, and applications for enhancing reasoning in large language models and large reasoning models since DeepSeek-R1.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.