When a language model is fine-tuned to follow instructions, some training answers contain facts the model never learned during pretraining. Fitting those answers teaches the model to state things it has no basis for, which is one documented source of hallucination, meaning fluent but false output. The usual remedies either delete those examples from the training set or reweight them. This proposal takes a different route. Keep the examples, but measure which layers of the network absorb most of the pressure to memorize unknown facts, and give those layers a smaller learning rate. Every other layer then stays free to learn format, style, and when to say "I don't know."
The claim is that this pressure is concentrated in a narrow band of layers, and that we can locate that band cheaply before and during training. The locator compares, layer by layer, the gradients from examples the model already knows with the gradients from examples it does not. A learning-rate schedule built from that comparison should cut hallucination and preserve the model's stored knowledge better than generic layer-wise schemes that spend the same total update budget. Generic schemes here means monotonic decay by depth, freezing the bottom layers, or importance scores tied to general competence. If the claim is right, we get two things: a practical optimizer-level tool that works even when unknown examples cannot be cleanly filtered out, and a map of where fine-tuning writes new facts inside a deep model.
What the literature establishes and leaves open
Two separate bodies of evidence motivate the idea. On the hallucination side, FLAME identifies "knowledge mismatch": supervised fine-tuning on facts absent from pretraining induces hallucination. FLAME fixes this at the data level by training only on answers the model can elicit from itself, and it applies uniform updates to all parameters. GRAIT reweights individual training samples so that learning to refuse does not become over-refusal. It also reports that the gradients for "I know" and "I don't know" examples are nearly orthogonal, with an inner product of about 0.008. Yet it still updates every layer with one learning rate.
Diagnostic work shows that truthfulness signals are not spread evenly across depth. INSIDE detects hallucination from middle-layer hidden states. Automatic Layer Selection for Hallucination Detection finds an intermediate-layer peak in intrinsic dimension that tracks semantic truth. Layer-wise Information Deficiency finds non-monotonic dips in the information passed between layers. HARP isolates a small subspace near the output that carries the useful signal. END shows that factual tokens gain probability sharply in late layers. All five are inference-time detectors or decoding rules; none changes how the model is trained. We use them as motivation, and two of them serve as measurement instruments after training.
On the optimization side, three papers supply the method family. ADEPT assigns asymmetric learning rates to attention and MLP sub-units, based on how important each unit is for general competence during continual pretraining. Intelligent Learning Rate Distribution finds, by black-box search on BERT, that good per-layer rates are non-monotonic. MetaLR and AutoLR learn per-layer rates with meta-learning or gradient-based automation. None of these targets factual hallucination.
Our proposal differs from ADEPT in the signal that sets the rates. ADEPT asks which units matter for general ability. We ask which units the unknown-fact examples specifically push on, relative to known-fact examples. This contrastive signal is the ingredient under test, and the ablations are designed to isolate it. The theoretical picture in Fine-Tuning Dynamics of In-Context Factual Recall, in which MLPs store facts and attention retrieves them, predicts that the signal should be dominated by middle-layer MLPs. We treat this as a hypothesis to check, not an assumption.
Core mechanism
First, a small labeled pool of training questions is split into "known" and "unknown" sets according to the base model's own sampled answers. The split follows the self-knowledge elicitation used by FLAME and GRAIT. For each layer (or each attention and MLP sub-unit) $l$, we compute $\bar g^{U}_l$ and $\bar g^{K}_l$, the mean fine-tuning gradients of that layer's parameters on the unknown and known subsets. From these we define the unknown-pressure share of the layer:
$$ s_l = \frac{\|\bar g^{U}_l\|^2}{\|\bar g^{U}_l\|^2 + \|\bar g^{K}_l\|^2}. $$
Each layer's learning rate is then set to
$$ \eta_l = \eta_0 \, \frac{L\, e^{-\beta (s_l - \bar s)}}{\sum_{j} e^{-\beta (s_j - \bar s)}}. $$
Here $\eta_0$ is the base learning rate, $L$ is the number of layers, $\bar s$ is the mean share across layers, and $\beta \ge 0$ controls how sharply the rates differ. The normalization keeps the mean per-layer rate at $\eta_0$. This matters for fairness: the method gets no advantage from simply updating less overall. It can only move plasticity between layers.
The shares are recomputed every few hundred steps on a small held-out pool. This addresses the coarse-interval limitation noted for ADEPT and costs a fraction of one epoch. When the training targets include refusals for unknown questions, as in refusal-aware tuning, the same rule should automatically keep late layers plastic if GRAIT's orthogonal refusal direction lives there. That localization is the second hypothesis we test.
Research Program
The work has three stages, and each one can fail informatively.
The first stage is a localization study, which needs no new training method. Using Llama-2-7B (the base model in GRAIT), Llama-3.1-8B, and Mistral-7B, we compute $s_l$ profiles on closed-book question-answering data from TriviaQA and Natural Questions. We ask three questions. Is the profile peaked, or flat? Does the peak coincide with the intermediate intrinsic-dimension peak and the middle layers where INSIDE's signal is strongest? Is the peak stable across models and data mixtures? A flat profile would falsify the core premise and end the method track. We would then report the negative result.
The second stage is controlled fine-tuning. We build supervised fine-tuning mixtures from TriviaQA and Natural Questions training questions, sweeping the fraction of unknown examples from 0 to 50 percent. This reproduces knowledge mismatch in a setting where we control the dose. We fine-tune with each learning-rate policy under a fixed step count, a fixed mean learning rate, and the same hyperparameter tuning budget for every policy.
The third stage transfers the method to realistic alignment. We fine-tune on long-form instruction data and evaluate on FLAME's Biography FActScore setup and Alpaca Eval. We also run refusal-aware tuning on the MMLU protocol used by GRAIT.
The measurements are held fixed across all stages. We record accuracy on held-out known questions, which tells us whether the model still states what it knew. We record hallucination rate on held-out unknown questions, meaning confident wrong answers rather than abstentions. For refusal-aware runs we report GRAIT's Truthful Helpfulness Score (THS) and error rate. We also report TruthfulQA and ARC-c as general checks.
As a mechanistic readout, we measure the detection accuracy (AUROC) of INSIDE's EigenScore and a HARP-style projection on the fine-tuned models. A policy that truly protects internal factual representations should keep these detectors working as well as they do on the base model. Uniform fine-tuning, if it damages those representations, should degrade them.
Baselines
The naive reference point is uniform full fine-tuning with a single tuned learning rate. It sets the floor for knowledge-mismatch damage, and every policy is compared to it at matched steps and mean rate.
The second group is the two cheap structural alternatives a practitioner would try first. One freezes the bottom half of the layers. The other is LoRA, which adds small low-rank weight updates, applied with one uniform rate. Each gets the same tuning budget as our method, meaning the same number of learning-rate trials. ADEPT reports gains of up to 5.58% on domain benchmarks and 5.76% on general benchmarks over full-parameter tuning in continual pretraining, but we found no reported result on Lacuna for either alternative on hallucination benchmarks. We therefore run both ourselves.
The third baseline is standard monotonic layer-wise decay. The rate is multiplied by a factor $\gamma<1$ per layer going downward, with $\gamma$ tuned, and the mean rate is renormalized to $\eta_0$. This is the strongest generic layer-wise competitor. Beating it shows that the shape of our profile matters, not just the fact of having different rates per layer.
The fourth and most important baseline is an ADEPT-style policy. It uses the same exponential mapping and the same normalization, but the scores come from general-competence importance instead of the known-versus-unknown contrast. This is the ablation that isolates our key ingredient. A second ablation keeps our exact set of rates but shuffles which layer receives which rate. That tests whether placement, rather than the spread of rates, drives any gain.
The strongest existing hallucination methods are data-level and sample-level. FLAME's self-knowledge fine-tuning reports a +5.6 point FActScore gain on Biography over standard fine-tuning plus preference optimization, with a 51.2% Alpaca Eval win rate. GRAIT reports a THS of 36.4 on MMLU, versus 34.0 for CRaFT and 20.4 for R-Tuning, and cuts the LLaMA2-7B MMLU error rate from 52.8% to 27.1%. We rerun both under our data and compute budget rather than comparing to the published numbers directly. We also test our schedule layered on top of each.
Beating a baseline means a lower unknown-question hallucination rate at equal or higher known-question accuracy and equal Alpaca Eval win rate, across three seeds with non-overlapping bootstrap intervals. For THS and FActScore, it means exceeding our reproduction of GRAIT and FLAME respectively. The published figures serve only as a sanity check on those reproductions.
Why not just…?
Why not just freeze the lower layers or use LoRA? Freezing assumes that knowledge sits at the bottom of the network and format sits at the top. The information dips reported in Layer-wise Information Deficiency and the intermediate truth peak found by intrinsic-dimension analysis both suggest that knowledge is concentrated in the middle, which bottom-freezing leaves fully plastic. LoRA restricts how much the model can change, not where the change happens. These are real competitors, though. If the localization study shows a peak in the lower layers, freezing could match our method. The matched-budget comparison in stage two will settle it, and we will report that outcome plainly.
Why not just use monotonic layer-wise decay? Black-box search in Intelligent Learning Rate Distribution found non-monotonic optimal profiles even for BERT classification, and the diagnostic papers point to a middle-layer peak that a monotonic schedule cannot represent. Still, if our measured $s_l$ turns out to be roughly monotonic, decay is a cheaper approximation to it. The tuned-decay baseline directly measures how much a non-monotonic shape adds.
Why not just filter out unknown examples, as FLAME does? This is the most serious threat. If unknown examples are removed perfectly, there is no mismatch left for layer rates to fix. The case for our method rests on filtering being imperfect in practice. Known and unknown are hard to label for long-form answers, and FLAME's pipeline depends on retrieval-based fact checking. Removing examples also discards the format supervision they carry. To test this honestly, we run FLAME-style filtering with deliberately noisy labels at several error rates, with and without our schedule. If filtering alone matches our method at realistic noise levels, we will conclude the method is only useful as a complement to filtering.
Why not just regularize toward the base model, for example with Elastic Weight Consolidation? This regularizer penalizes changes to parameters that were important for earlier tasks. It protects everything it deems important, not specifically the stored facts. The layer-rate search in Intelligent Learning Rate Distribution beat it by about 5% under dataset shift. We include a regularization toward the base model as an extra baseline only in stage two, where it is cheap to run.
Why not just detect hallucinations at test time? Detectors such as INSIDE need 10 to 20 sampled answers per query. END's decoding rule can only amplify knowledge that is already present and cannot repair damage done during fine-tuning. We treat these methods as complementary: in our design they measure whether the stored knowledge survived training.
Decisive checks
The idea rests on three results. First, the $s_l$ profile must be peaked and reproducible across the three models. If it is flat, the premise fails. Second, in the dose-response experiment, the gap between our method and the ADEPT-style ablation must widen as the unknown fraction grows. If it does not, the contrastive signal is not what drives any gain. Third, the fine-tuned models' EigenScore and HARP detection accuracy must stay closer to base-model levels under our schedule than under uniform tuning. That would show the method protects internal factual representations rather than merely changing output style. A positive result on all three, plus a gain over noisy-label filtering, would support the claim that fine-tuning writes new facts into a locatable band of layers, and that slowing learning there is an effective defense against the hallucination that fine-tuning induces.
References
- FLAME : Factuality-Aware Alignment for Large Language Models
- GRAIT: Gradient-Driven Refusal-Aware Instruction Tuning for Effective Hallucination Mitigation
- INSIDE: LLMs' Internal States Retain the Power of Hallucination Detection
- Automatic Layer Selection for Hallucination Detection
- Detecting LLM Hallucination Through Layer-wise Information Deficiency: Analysis of Ambiguous Prompts and Unanswerable Questions
- HARP: Hallucination Detection via Reasoning Subspace Projection
- Improve Decoding Factuality by Token-wise Cross Layer Entropy of Large Language Models
- ADEPT: Continual Pretraining via Adaptive Expansion and Dynamic Decoupled Tuning
- Intelligent Learning Rate Distribution to reduce Catastrophic Forgetting in Transformers
- MetaLR: Layer-wise Learning Rate based on Meta-Learning for Adaptively Fine-tuning Medical Pre-trained Models
- Layerwise Change of Knowledge in Neural Networks
- Fine-Tuning Dynamics of In-Context Factual Recall in Transformers