One-line positioning: A Johns Hopkins team proposes Neural Trust Functions (NTF), which use the weak teacher's final-layer hidden states—not output-layer confidence—to judge whether weak labels are reliable. Across three task families (world knowledge, math reasoning, chess), strong students trained on filtered weak supervision nearly match or exceed ground-truth-trained students, with 104-110% "super-recovery" in some settings. Even more striking, iterative chained training compounds gains like a snowball: Qwen3-14B reaches 48.2% on chess, far above 40.0% with ground-truth training. ICML 2026.
---
1. The Problem: Why Weak-to-Strong Generalization Hits a Ceiling
Burns et al.'s 2023 paper launched weak-to-strong generalization research: use a large model as student and a small model as teacher, aiming to let the strong student surpass the weak teacher without ground-truth labels. But nearly all follow-up work hit a ceiling—students trained on weak supervision never fully catch up to ground-truth-trained students.
Two causes are typically cited: 1. Error propagation: the weak teacher's wrong labels are inherited unless the data geometry happens to let the strong model correct them 2. Representation gap: the weak teacher's representation space misses task-relevant directions, so training signal is missing
Prior solutions focus on output-level heuristics: entropy, self-consistency, multi-model agreement, verbalized confidence. But these signals are often poorly calibrated on complex tasks—scoring "confident wrong answers" high and "correct but uncertain solutions" low. This is especially fatal under distribution shift.
---
2. Core Insight: Internal Representations Know When the Answer Is Correct, Even When the Output Is Wrong
The paper's breakthrough builds on Kadavath et al. (2022): a large model's intermediate representations contain separable signals of answer correctness that get masked in the final decoded output.
So the authors reframe weak-to-strong generalization as a data selection problem: not every weak label deserves trust; the key is identifying which ones are reliable enough to serve as training signal.
They propose Trust Functions—a scalar trust score between 0 and 1 for each weak label, then train the strong student on high-scoring labels.
---
3. How Neural Trust Functions Work
Input: Final-Layer Hidden State
NTF never touches output probabilities. Its input is the final-layer hidden state when the weak teacher generates the final answer token. That token's attention has seen the full input and all intermediate reasoning, making it a compact "correctness summary vector."
Architecture: A Small Residual MLP
NTF is just a small MLP with residual connections—RMSNorm-SwiGLU blocks + Dropout + stochastic depth, ending in a linear head whose logits pass through a sigmoid to produce 0-1.
Cost is near zero. The teacher's forward pass is needed anyway to generate answers; hidden states come for free. NTF training and inference cost far less than the teacher's forward pass—pipeline time is dominated by teacher inference.
Training: Binary Cross-Entropy
NTF is trained on a labeled source distribution, constructing binary labels (correct/incorrect) by comparing the weak teacher's predictions with ground truth, using class-reweighted BCE to handle imbalance.
---
4. Zero-Shot Cross-Distribution Scoring
A key design: NTF is trained only on a labeled source distribution, then deployed zero-shot on an unlabeled target distribution—provided both share the same domain and task interface (e.g., both multiple-choice, or both math).
This is highly practical: you often have labeled old data (e.g., MATH) but unlabeled new targets (e.g., AIME competition problems). NTF learns "how to judge weak label reliability" on MATH and directly filters weak labels on AIME.
---
5. Results Across Three Task Families: Nearly Lossless, Sometimes Better
Metric: Recovery
Base = untrained baseline, GT = ground-truth-trained. 100% means fully matching GT; >100% means exceeding it.
1. World Knowledge (MCQA)
| Teacher | Student | NTF | GT | Recovery | |---------|---------|-----|----|----------| | OLMo2-1B | OLMo2-7B | 73.7 | 73.8 | 98.9% | | OLMo2-1B | OLMo2-13B | 80.9 | 81.2 | 95.9% | | Qwen3-0.6B | Qwen3-1.7B | 75.0 | 74.8 | 103.5% (super-recovery) | | Qwen3-0.6B | Qwen3-14B | 87.1 | 87.0 | 101.3% (super-recovery) |
Of 8 settings, 5 show no significant difference from GT, 1 significantly exceeds GT. NTF beats output-level baselines (I-Confidence, Ensemble, Reward Model) across the board.
2. Math Reasoning (GRPO Training)
| Teacher | Student | NTF | GT | Recovery | |---------|---------|-----|----|----------| | Qwen3-1.7B | Qwen3-4B | 22.0 | 22.9 | 92.5% | | Qwen3-4B | Qwen3-8B | 26.6 | 27.4 | 91.0% | | Qwen3-8B | Qwen3-8B | 27.9 | 28.7 | 91.7% | | Qwen3-8B | Qwen3-4B | 27.4 | 28.4 | 89.2% |
All settings reach 89-92% recovery, half statistically indistinguishable from GT. The teacher itself may have <5% accuracy on the target distribution (Qwen3-1.7B on AIME), yet NTF still extracts a high-purity subset from its weak labels.
3. Strategy Games (Chess Puzzles)
The most striking domain. Chess has an external evaluator (Stockfish) that objectively grades every move.
| Teacher | Student | NTF | GT | Recovery | |---------|---------|-----|----|----------| | Qwen3-0.6B | Qwen3-1.7B | 15.5 | 14.9 | 104.4% (super-recovery) | | Qwen3-0.6B | Qwen3-14B | 44.1 | 39.9 | 110.4% (super-recovery) | | OLMo2-1B | OLMo2-1B | 37.4 | 37.7 | 99.2% | | OLMo2-1B | OLMo2-13B | 41.5 | 54.5 | 76.1% (OLMo2 family less stable in this domain) |
The Qwen3 family excels at chess: 4 super-recovery settings, 1 near-lossless. NTF doesn't just match ground truth—it surpasses it.
---
6. The Snowball Effect: Iterative Chained Weak-to-Strong Training
The most counterintuitive part: if NTF-filtered weak supervision rivals ground truth, the trained student can itself become the next round's teacher—a weak-to-strong chain.
Starting from the weakest Qwen3-0.6B, iterating round by round:
| Method | Qwen3-4B | Qwen3-8B | Qwen3-14B | |--------|---------|---------|----------| | Naive Shallow | 27.0 | 33.7 | 38.1 | | Naive Chain | 30.1 | 34.2 | 39.1 | | NTF Shallow (0.6B) | 35.4 | 38.0 | 44.1 | | NTF Shallow (8B) | — | — | 46.1 | | NTF Chain | 36.9 | 40.1 | 48.2 | | Ground Truth | 36.2 | 37.0 | 40.0 |
NTF Chain's Qwen3-14B hits 48.2%, 8.2 points above ground truth's 40.0%. Chained NTF simultaneously beats:
- Shallow NTF transferred directly from the weakest teacher
- Shallow NTF transferred from the strongest available weak teacher (8B)
- Naive chained baselines (no NTF filtering)
- Ground-truth label training
- The distribution has substantial negative mass (NTF moves stronger than ground-truth moves)
- 66.1% of retained moves lead to mate
- Uzunoğlu, A., Zhang, A., & Khashabi, D. (2026). *Trust Functions: Near-Lossless Weak-to-Strong Generalization by Learning When to Trust the Weak Teacher*. arXiv:2606.01000. ICML 2026. https://arxiv.org/abs/2606.01000
- Burns, C., et al. (2023). *Weak-to-Strong Generalization: Eliciting Strong Capabilities with Weak Supervision*. arXiv:2312.09390
- Kadavath, S., et al. (2022). *Language Models (Mostly) Know What They Know*. arXiv:2207.05221
Implication: with weak labels + trust filtering + iteration, the final result can systematically exceed ground-truth training. Label quality isn't the only bottleneck—label selection strategy matters just as much.
---
7. Why Can NTF Beat Ground Truth? Three Mechanisms
The authors dissect chess in depth, since Stockfish objectively grades every move.
Mechanism 1: Conservative Filtering Forms an Implicit Easy-First Curriculum
NTF prefers low-difficulty (low-rated) puzzles. This isn't a bug—it's a feature: easy examples form an implicit easy-to-hard curriculum that helps smaller models. But ablations show difficulty matching explains only part of the gain and vanishes on larger models, so it's not the whole story.
Mechanism 2: 66.1% of "False Positives" Are Actually Stronger Than Ground Truth
Among NTF-retained "false positives" (moves NTF trusts but the dataset labels wrong), Stockfish evaluation shows:
Many answers labeled "wrong" by the dataset are actually better than the labeled ground truth. NTF's "errors" often uncover dataset labeling errors.
Experimental check: relabeling NTF-retained instances with ground truth (NTF-GT) performs slightly *worse* than NTF itself—those "wrong" weak labels genuinely carry useful supervision.
Mechanism 3: More Coherent Gradient Directions
Gradients from NTF-filtered data concentrate more energy in the dominant subspace, with faster singular-value decay. Updates are more "consistent"—samples push the model in similar directions instead of pulling against each other. This explains why fewer, higher-quality samples can beat full ground-truth training.
---
8. Controllable Data Selection
In deployment, you don't know how many samples to keep. The authors provide a Hoeffding-inequality-based calibration procedure:
Using a small labeled slice of target data, compute an upper confidence bound U(θ) on the noise rate for each candidate trust threshold θ, and pick the most inclusive threshold satisfying U(θ) ≤ α.
Experiments: the selected θ* = 0.895 retains 16.1% of the deployment pool, with actual noise rate below the target α = 0.1—a tuning-free tool for practitioners to control label quality.
---
9. Limitations and Boundaries
1. Needs labeled source data: NTF requires some labeled data to train; it cannot start from fully zero labels (though existing labeled legacy data works) 2. Outcome supervision only: currently predicts only final-answer correctness; process supervision (per-step reasoning rewards) is unexplored 3. Architecture limits: only a simple MLP on a single-token hidden state; more complex architectures (e.g., attention-based NTF) could capture more temporal reasoning signal 4. Cross-domain failure: zero-shot transfer works within the same task interface (same domain), but degrades across task interfaces (e.g., MCQA → chess)
---
10. Significance: A Paradigm Shift in Data Selection
The paper's core contribution is not another training trick but a reframing of weak-to-strong generalization—from "how to train models to correct weak labels" to "how to filter the trustworthy subset of weak labels."
It demonstrates that: 1. A weak teacher's internal representations know when it is bluffing better than its output probabilities do 2. The right data matters more than right-labeled data volume—NTF's retained 16% high-quality subset outperforms full ground truth 3. Weak-to-strong can be an iterative snowball, not a one-shot leap—each round's student becomes the next teacher, compounding gains
For resource-constrained teams, vertical domains with expensive annotation (medicine, science), and an era flooded with mixed-quality synthetic data, NTF offers a near-zero-cost filtering mechanism.
---