One-line positioning: A Johns Hopkins team proposes Neural Trust Functions (NTF), which use the weak teacher's last-layer hidden state — not output-layer confidence — to judge whether a weak label is reliable. Across three task families (world knowledge, math reasoning, chess), strong students trained on NTF-filtered weak supervision nearly match or exceed ground-truth training, with some settings achieving 104-110% "super-recovery." Even more striking, iterative chained training compounds gains like a snowball: a final Qwen3-14B reaches 48.2% on chess, far above the 40.0% achieved with ground-truth labels. ICML 2026.
---
1. The Problem: Why Weak-to-Strong Generalization Hits a Ceiling
Burns et al. (2023) opened the study of weak-to-strong generalization: use a large model as the student and a small model as the 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 quite matched students trained on ground-truth labels.
Two causes are commonly cited: 1. Error propagation: the weak teacher's wrong labels are inherited by the student 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 absent.
Prior solutions rely on output-level heuristics: entropy, self-consistency, multi-model agreement, verbalized confidence. These signals are often poorly calibrated on complex tasks — scoring "confident wrong answers" high and "correct but uncertain solutions" low. This is especially damaging under distribution shift.
---
2. Core Insight: Internal Representations Know the Answer Is Right Even When the Output Is Wrong
The paper's breakthrough builds on a finding by Kadavath et al. (2022): a large model's intermediate-layer representations contain separable signals about answer correctness, but these signals are masked in the final decoded output.
So the authors reframe the problem: weak-to-strong generalization is fundamentally a data selection problem. Not every weak label deserves trust; the key is finding which weak labels are reliable enough to serve as training signal.
They propose Trust Functions — a scalar trust score between 0 and 1 for each weak label — and train the strong student only on high-score labels.
---
3. How Neural Trust Functions (NTF) Work
Input feature: the last-layer hidden state
NTF never touches output-layer probabilities. Its input is the last-layer hidden state of the weak teacher at 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 a 0-1 score.
Computational 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, so pipeline time is dominated by teacher inference.
Training: binary cross-entropy
NTF is trained on a labeled source distribution by comparing teacher predictions against ground truth to construct binary labels (correct/incorrect), using class-reweighted binary cross-entropy to handle class imbalance.
---
4. Zero-Shot Cross-Distribution Deployment
A key design: NTF only needs training on a labeled source distribution, then deploys zero-shot on an unlabeled target distribution — as long as both share the same domain and task interface (e.g., both multiple-choice, both math). In practice: you have labels on old data (e.g., MATH) but not on new target data (e.g., AIME competition problems). NTF learns "how to judge weak labels" on MATH and directly filters weak labels on AIME.
---
5. Results Across Three Task Families
Metric: Recovery
Base = untrained baseline, GT = ground-truth training. 100% means matching ground truth; >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) | | Qwen3-0.6B | Qwen3-14B | 87.1 | 87.0 | 101.3% (super) |
Of 8 settings, 5 are statistically indistinguishable from GT; 1 significantly exceeds GT. NTF outperforms 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 indistinguishable from GT. Notably, the teacher itself may score <5% 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 precisely grades each move.
| Teacher | Student | NTF | GT | Recovery | |---------|---------|-----|----|----------| | Qwen3-0.6B | Qwen3-1.7B | 15.5 | 14.9 | 104.4% (super) | | Qwen3-0.6B | Qwen3-14B | 44.1 | 39.9 | 110.4% (super) | | OLMo2-1B | OLMo2-1B | 37.4 | 37.7 | 99.2% | | OLMo2-1B | OLMo2-13B | 41.5 | 54.5 | 76.1% (OLMo2 family less stable here) |
The Qwen3 family excels at chess: 4 settings super-recover, 1 near-lossless. NTF doesn't just match ground truth — it beats it.
---
6. The Snowball Effect: Iterative Chained Weak-to-Strong Training
The most counterintuitive part: if NTF-filtered weak supervision already rivals ground truth, the trained student can become the next round's teacher — forming a weak-to-strong chain.
Starting from the weakest Qwen3-0.6B and iterating:
| 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 |
The final NTF-chained Qwen3-14B hits 48.2%, 8.2 points above ground-truth training's 40.0%. Chained NTF beats shallow NTF from the weakest teacher, shallow NTF from the strongest available teacher (8B), naive chaining without filtering, and ground-truth training itself.
Conclusion: with weak labels + trust filtering, iterating can systematically surpass ground-truth training. Label quality is not 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 creates an implicit easy-first curriculum
NTF prefers low-difficulty (low-rated) puzzles. This is a feature, not a bug — easy examples form an implicit "easy-first" curriculum that helps smaller models. But ablations show difficulty matching only explains part of the gain and vanishes for larger models, so it's not the whole story.
Mechanism 2: 66.1% of "false positives" are actually stronger than ground truth
Among the "false positives" NTF retains (high-trust moves labeled wrong in the data), Stockfish evaluation shows:
- The distribution has substantial negative values (NTF's moves are stronger than ground-truth moves)
- 66.1% of retained NTF moves lead to a forced win (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
In other words, many answers the dataset marks as "wrong" are actually better than the labeled ground truth. NTF's "errors" are often discoveries of dataset labeling errors.
Experimental check: relabeling NTF-retained instances with ground truth (NTF-GT) performs slightly *worse* than NTF itself — confirming those "wrong" weak labels carry useful supervision.
Mechanism 3: More consistent gradient directions
Gradients from NTF-filtered data concentrate more energy in the dominant subspace, with faster singular-value decay. Updates point in more "consistent" directions — samples push the model the same way instead of pulling apart. This explains why NTF's smaller, higher-quality subset can outperform training on all ground-truth samples.
---
8. Risk-Controlled 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 holdout from the target data, compute an upper confidence bound U(θ) on the noise rate for each candidate trust threshold θ, and pick the largest inclusion threshold satisfying U(θ) ≤ α.
In experiments, the chosen θ* = 0.895 kept 16.1% of the deployment pool with actual noise below the target α = 0.1 — a tuning-free tool for controlling label quality.
---
9. Limitations
1. Requires labeled source data: NTF needs some labeled data to train; it cannot start from a fully zero-label setting (though existing labeled old data suffices). 2. Outcome supervision only: it predicts final-answer correctness only; process supervision (e.g., per-step reasoning rewards) is unexplored. 3. Architecture limits: only a simple MLP on a single-token hidden state; richer architectures (e.g., attention-based NTFs) might capture more sequential 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 redefinition 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 understand when it is wrong better than its output probabilities do. 2. The right data matters more than correct data — NTF's retained 16% high-quality subset outperforms the full ground-truth set. 3. Weak-to-strong need not be a one-shot jump but can be an iterative snowball — each round's student becomes the next round's teacher, compounding gains.
For resource-constrained teams, vertical domains (medicine, science) with expensive labeling, and an era awash in uneven synthetic data, NTF offers a near-zero-cost filtering mechanism.
---