REVIEW 60 cited by
Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision
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.
Forward citations
Cited by 60 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.
-
Bridging Compute- and Data-Optimal Pretraining
Pretraining loss obeys a single law in which repeated or paraphrased tokens count as η(N, data-per-parameter, expansion-ratio) fresh tokens, with total effective data saturating as derived tokens grow.
-
Understanding Reasoning from Pretraining to Post-Training
A joint scaling law: post-RL chess and math performance is predictable from pretraining loss, RL improvement rate grows with pretraining tokens, and RL both amplifies and discovers moves depending on difficulty.
-
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.
-
Search Beyond What Can Be Taught: Evolving the Knowledge Boundary in Agentic Visual Generation
A generator-specific knowledge boundary is discoverable by co-training: DPO teaches what can be internalized, then rejection finetuning calibrates search to only what remains external.
-
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 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.
-
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.
-
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.
-
A Scalable Approach to Evaluating Moral Sensitivity in LLMs
Under morally irrelevant noise, eight LLMs preserve the semantic content of identified moral features above calibrated floors, despite significant changes in feature counts.
-
Weak-to-Strong On-Policy Distillation
A strong LLM is improved by distilling from the logit difference of two weaker models instead of from a stronger teacher.
-
Shared SFT Lessons Across Alignment, Model Organisms, and Toy Models
SFT lessons — reason-based training, on-model replay, and wash-out robustness — transfer across toy models, model organisms, and alignment SFT, improving the capability–safety tradeoff.
-
When Do Agent Loops Mistake Stagnation for Progress? Self-Evaluation Bias and Externally Grounded Verification in Long-Running Autonomous LLM Agent Loops
For open-ended agent goals whose success lives outside the transcript, even a strong in-band judge fails; out-of-band world-state gating is structurally required to stop the progress mirage.
-
Weak-to-Strong Generalization via Direct On-Policy Distillation
Transferring a weak model’s RL-induced log-ratio policy shift on a strong student’s own rollouts raises AIME accuracy more cheaply than imitating the weak teacher or running matched-step RL on the student.
-
Search Beyond What Can Be Taught: Evolving the Knowledge Boundary in Agentic Visual Generation
Selective search plus generator-reasoner co-training improves knowledge-grounded image generation, but the reported gains are scored by the same VLM judge used to train the system.
-
Search Beyond What Can Be Taught: Evolving the Knowledge Boundary in Agentic Visual Generation
A co-training framework (DPO on the generator, then rejection finetuning on the reasoner) discovers a generator-specific knowledge boundary that enables selective search for world-knowledge-grounded image generation, ...
-
Multi-Turn On-Policy Distillation with Prefix Replay
ReOPD offline-distills multi-turn agentic LLMs via teacher-prefix replay plus step-decay sampling, matching online OPD accuracy at ≥4× speed with zero tool calls.
-
LivingArena: Do LLMs Know What Other LLMs Don't? Peer-Probing as Scalable Evaluation
Large language models can probe each other's weaknesses in a self-running tournament, producing a stable Elo leaderboard without a static question bank.
-
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.
-
DenoiseRL: Bootstrapping Reasoning Models to Recover from Noisy Prefixes
DenoiseRL injects wrong reasoning prefixes from weak models into RL rollouts and trains the policy to recover, improving GRPO/DAPO mathematical reasoning on named 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
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.
-
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...
-
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
Automating alignment research with AI agents risks generating hard-to-detect errors in fuzzy tasks, producing misleading safety evaluations even without deliberate sabotage.
-
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.
-
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 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.
-
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 ...
-
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.
-
CPMobius: Iterative Coach-Player Reasoning for Data-Free Reinforcement Learning
A cooperative Coach–Player RL loop, rewarded by Player validation-progress, lifts math reasoning by 2–5 points while the Player trains only on self-generated pseudo-labeled tasks.
-
CPMobius: Iterative Coach-Player Reasoning for Data-Free Reinforcement Learning
CPMobius uses iterative coach-player reinforcement learning to improve mathematical reasoning in LLMs without external training data, yielding +4.9 average accuracy gains on Qwen2.5-Math-7B-Instruct.
-
When and Why Does Multi-Agent Debate Fail and Does It Really Underperform?
A collaborative, non-zero-sum debate protocol (ColMAD) outperforms competitive multi-agent debate and modestly beats single LLMs at detecting errors in model responses.
-
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.
-
ACE and Diverse Generalization via Selective Disagreement
ACE learns an ensemble of classifiers that agree on labeled data but confidently and selectively disagree on target-distribution data, recovering diverse human-interpretable concepts under complete spurious correlation.
-
ReCode: Improving LLM-based Code Repair with Fine-Grained Retrieval-Augmented Generation
ReCode improves LLM-based code repair by combining algorithm-type prediction with dual-encoder retrieval of authentic bug-fix pairs, and introduces RACodeBench for evaluation.
-
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.
-
Code Monitor Red Teaming for Public-Test-Passing Code
Weak LLM verifiers miss most hidden bugs in public-test-passing code, and adversarial pressure makes that monitoring worse.
-
Proxy Exploration and Reusable Guidance: A Modular LLM Post-Training Paradigm via Proxy-Guided Update Signals
Relative policy-improvement signals from a weak proxy model, after simple calibration, can be transferred to improve stronger primary LLMs without re-exploring on the primary.
-
Multi-Turn On-Policy Distillation with Prefix Replay
By replaying teacher prefixes with a step-decay schedule, multi-turn on-policy distillation can run without live environment interaction, matching or slightly beating online OPD accuracy.
-
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.
-
OpenEvoShield: Dual Non-Stationary Continual Defense for Open-World Multi-Agent System Attacks
OpenEvoShield claims to defend LLM multi-agent systems against evolving and novel attacks by combining asymmetric-rate continual learning with energy-based OOD detection.
-
Honest Reporting in Scored Oversight: True-KL0 Property via the Prekopa Principle
Conditional on an ad-hoc structural assumption about the reporting family, the paper claims and numerically bounds the incentive ratio R ≤ 0.7681 for interface dimension d ≤ 4, asserting that honest reporting dominate...
-
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.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.