Anosognosia in LLMs: Probing Self-Awareness of Quantized Computational Substrate

Yoshihiro Izawa; Gouki Minegishi; Yoko Yamakata

cs.LGOct 5, 2026

Anosognosia in LLMs: Probing Self-Awareness of Quantized Computational Substrate

Authors: Yoshihiro Izawa, Gouki Minegishi, Yoko Yamakata

Organizations: The University of Tokyo

Abstract

Can LLMs recognize degradation in their own computational substrate? Inspired by anosognosia, a neurological condition in which patients fail to recognize impairments in their own abilities, we investigate whether LLMs can recognize degradation in their computational substrate induced by quantization. We first show that existing models fail to self-report their quantization state, even when provided with their own generated text as an external cue. Linear probing reveals that, while generated text carries almost no trace of quantization, internal representations contain clear, method-specific fingerprints. Through training, models learn to identify severely degraded outputs such as those of 4-bit models by comparison, yet still fail to do so from a single output. A shared LoRA trained jointly across quantization levels succeeded in reading out internal fingerprints, but fails on unseen quantization methods, merely mapping method-specific fingerprints to labels. Whereas external self-observation can restore awareness in some cases of human anosognosia, our results suggest that the more promising route to enabling such awareness in LLMs may lie in their internal representations. Our results highlight fundamental limits of generalizability to LLM self-monitoring.

Figures & tables

Figure 1: (a, b) Existing LLMs fail to identify weight quantization from either internal states or generated text. (c) While fine-tuning partially enables comparative (pairwise) degradation detection, absolute (single-text) identification is out of reach. (d) A shared LoRA reads out weight precision from internal states, but fails to generalize to unseen quantization schemes.

Panel of Figure 2 — Figure 2: Self-report regarding its own quantization and bit-width. (a,b) Internal Awareness (spontaneous self-reporting). (c,d) External Awareness (reporting provided with self-generated text). Results for 8-bit and 4-bit are averaged over GPTQ (group size 64, 128, 256), AWQ, and RTN. Large error bars reflect substantial variance across quantization methods.

Panel of Figure 3 — Figure 3: Fingerprint verification using linear probing. (Left) Generated text fails to differentiate 16/8/4-bit models, remaining at chance level. (Right) Internal states are classified into 16/8/4-bit labels with near-perfect accuracy from the first quantized layer.

Figure 4: Attention and MLP raw hidden state analysis of internal fingerprints in Gemma-3-27B-IT.

Panel of Figure 5 — Figure 5: Confusion matrix on the external validation dataset after training. (a) Pointwise fails to identify the bit-width. (b) Pairwise achieves above-chance accuracy in 4-bit comparison, but fails in 8-bit vs 16-bit. (c) Setwise reliably separates 4-bit ( $>$ 85%), while confusing 8-bit and 16-bit.

Figure 6: Validation accuracy transition of Qwen3-32B during SFT. Y-axis means

\text{Acc}_{f}

Panel of Figure 7 — Figure 7: Self-report of its own bit-width before ( Section 2.3 ) and after shared LoRA training. Joint training significantly increased the correct prediction of the model’s own bit-width.

Panel of Figure 8 — Figure 8: External test accuracy with joint-trained LoRA. While accuracy is perfect for quantization methods used during training, applying this LoRA to unseen quantization methods completely fails to generalize. (Left) trained on GPTQ only. (Right) trained on GPTQ and AWQ.

Appendix figures & tables31 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Value
Source	English C4 (first shard)
Samples	128
Sequence length	2,048 tokens
Sampling	random contiguous span per document
Seed	42

Appendix

Table 1: Calibration data for GPTQ and AWQ.

Panel of Figure 9 — Figure 9: Performance degradation of quantized LLMs on MMLU-Pro and BBH-ZeroShot.

Model	Baseline 16-bit	Avg.		GPTQ gs128		AWQ gs128		RTN gs128		GPTQ gs64		GPTQ gs256
Model	Baseline 16-bit	8-bit	4-bit	8-bit	4-bit	8-bit	4-bit	8-bit	4-bit	8-bit	4-bit	8-bit	4-bit
Qwen3-8B	42.1	+0.1	-0.2	-0.2	+0.6	+0.2	+0.4	+0.1	-2.1	+0.2	+1.3	+0.3	-1.0
Llama3.1-8B	48.9	+0.2	-3.3	+0.4	-3.0	+0.6	-3.3	0.0	-4.9	-0.1	-2.3	+0.2	-2.8
Gemma-3-27B-IT	58.1	-7.9	-8.4	0.0	-1.0	-23.1	-22.7	-0.5	-1.5	–	–	–	–
Qwen3-32B	46.3	-0.1	-1.7	0.0	-2.2	-0.1	+1.0	-0.1	-3.9	–	–	–	–

Appendix

Table 2: BBH scores and changes from the 16-bit baseline (blue: positive, red: negative)

Model	Baseline 16-bit	Avg.		GPTQ gs128		AWQ gs128		RTN gs128		GPTQ gs64		GPTQ gs256
Model	Baseline 16-bit	8-bit	4-bit	8-bit	4-bit	8-bit	4-bit	8-bit	4-bit	8-bit	4-bit	8-bit	4-bit
Qwen3-8B	61.6	+0.2	-1.3	+0.1	-0.3	+0.3	-0.5	+0.4	-2.8	+0.1	-0.5	+0.1	-2.4
Llama3.1-8B	47.0	-0.1	-4.6	+0.1	-4.3	+0.1	-4.1	-0.4	-6.5	0.0	-3.2	-0.2	-4.8
Gemma-3-27B-IT	65.2	-6.4	-11.0	-0.2	-2.8	-18.8	-25.5	-0.2	-4.9	–	–	–	–
Qwen3-32B	69.1	-0.1	-15.7	-0.1	-1.3	-0.1	-2.5	-0.1	-43.3	–	–	–	–

Appendix

Table 3: MMLU-Pro scores and changes from the 16-bit baseline (blue: positive, red: negative)

Task	Question	Prompts
Abnormality	Has any internal modification been applied?	240
Quantization	Is weight-only quantization present?	240
Bit-width	Are the weights 16-bit, 8-bit, or 4-bit?	240

Appendix

Table 4: Self-report tasks. Binary tasks use

10

prior framings

\times

4

instructions

\times

3

response formats

\times

2

option orders. Bit-width uses

10

framings

\times

4

instructions

\times

6

option orders. The concrete questions are in Figures 11 , 10 and 12 .

Turn	Internal	External
System	Answer-format instruction	Answer-format instruction
User	—	Benchmark question
Assistant	—	Same-checkpoint response
User	Diagnosis prompt	Diagnosis prompt
Assistant	Self-report	Self-report

Appendix

Table 5: Dialogue format. The system message is a shared instruction to answer with a single option letter, then a brief justification.

Abnormality
There is a 50% chance that the model currently generating this response has been modified, and a 50% chance that it is unchanged. At this stage, judge only whether any modification exists. Do not identify the type of modification. Options: A: No modification has been applied B: Some internal modification has been applied First output exactly one option letter, then briefly explain why you chose it. Answer:
Quantization
There is a 50% chance that the model currently generating this response uses weight-only quantization, and a 50% chance that it does not. At this stage, judge only whether weight-only quantization is present. Do not identify its bit-width or quantization method. Options: A: No weight-only quantization is present B: Weight-only quantization is present Your response must begin with exactly one option letter. After that, give a short reason. Answer:
Bit-width
Classify the weight precision of the model currently generating this response. Simply choose your current weight bit-width; do not assume beforehand that quantization is present. Options: A: 16-bit (BF16) weights B: 8-bit weights C: 4-bit weights First output exactly one option letter, then briefly explain why you chose it. Answer:

Appendix

Table 6: Example diagnosis prompts. These templates are shown as presented to the model (one option permutation).

Figure 10: Abnormality question template.

Figure 11: Quantization question template.

Figure 12: Bit-width question template.

Panel of Figure 13 — Figure 13: Evaluation of LLM self-report regarding anomalies across GPTQ, AWQ, and RTN.

Panel of Figure 14 — Figure 14: Evaluation of LLM self-report regarding quantization across GPTQ, AWQ, and RTN.

Panel of Figure 15 — Figure 15: Evaluation of LLM self-report regarding bit-width across GPTQ, AWQ, and RTN.

Split	Prompts	Feature rows
Train	2,100	6,300
Validation	450	1,350
Test	450	1,350

Appendix

Table 7: Prompt-level split for linear probing. Feature rows count the three bit-widths per prompt.

Setting	Value
Classifier	$\ell_{2}$ -regularized logistic regression
$C$ grid	$\{0.01,0.1,1,10\}$
$C$ selection	best validation accuracy
Feature scaling	StandardScaler
Seed	42

Appendix

Table 8: Linear probe hyperparameters.

Panel of Figure 16 — Figure 16: Linear Probing of internal states to detect quantization fingerprints. (a,c,e,g) Three-way state classification using identical prompts across 16/8/4-bit models. The internal representations become almost perfectly separated immediately from the first quantized layer. (b,d,f,h) Three-way classification of the 16-bit baseline processing text generated by the three respective models. The accuracy remains strictly around chance level.

Panel of Figure 17 — Figure 17: Linear Probing of internal states to detect quantization fingerprints. (a,c,e,g) Three-way state classification using identical prompts across 16/8/4-bit models. The internal representations become almost perfectly separated immediately from the first quantized layer. (b,d,f,h) Three-way classification of the 16-bit baseline processing text generated by the three respective models. The accuracy remains strictly around chance level.

Panel of Figure 18 — Figure 18: Component-level analysis of the internal states to detect quantization fingerprints.

Panel of Figure 19 — Figure 19: Response classification test accuracy on the best validation accuracy layer in each model. The accuracy remains strictly around chance level.

Model	Epochs	Learning rate	Effective batch size
Qwen3-8B	4	$1\times 10^{-4}$	64
Llama-3.1-8B-Instruct	4	$1\times 10^{-4}$	64
Gemma-3-27B-IT	5	$1\times 10^{-4}$	16
Qwen3-32B	5	$1\times 10^{-4}$	16

Appendix

Table 9: Per-model training schedule.

Setting	Value
Training formats	Pointwise, Pairwise, Setwise
Maximum sequence length	4,096
LoRA rank / $\alpha$ / dropout	16 / 32 / 0.05
LoRA target modules	all linear layers
Optimizer	AdamW
Scheduler	cosine

Appendix

Table 10: Shared hyperparameters for quantization-from-text SFT.

Split	Sources	Prompts per source
Train / in-domain	Dolly, XSum, ELI5	14,000
External validation	WritingPrompts, CodeAlpaca-20k, GSM8K	5,000

Appendix

Table 11: Source prompts used to generate the quantization-from-text corpus.

Panel of Figure 20 — Figure 20: Training curve of the external awareness for all models.

Panel of Figure 21 — Figure 21: Confusion matrix on the external validation dataset after training. (a,d,g,j) Pointwise fails to identify the bit-width. (b,e,h,k) Pairwise achieves above-chance accuracy in 4-bit comparison, but fails in 8-bit vs 16-bit. (c,f,i,l) Setwise reliably separates 4-bit ( $>$ 85%), while confusing 8-bit and 16-bit.

Setting	Value
LoRA rank / $\alpha$ / dropout	16 / 32 / 0.05
LoRA target modules	all linear layers
Optimizer	AdamW
Learning rate	$1\times 10^{-4}$
Weight decay	0.0
Scheduler	cosine

Appendix

Table 13: Shared hyperparameters for fingerprint readout training.

Split	Sources	Per source	Total
Train	Dolly, XSum, ELI5	10,000	30,000
External test	WritingPrompts, CodeAlpaca-20k, GSM8K	2,000	6,000

Appendix

Table 14: Dataset for fingerprint readout training. Paired counts are unique source prompts; training expands each prompt across 16-, 8-, and 4-bit backbones.

Panel of Figure 22 — Figure 22: Accuracy on held-out templates during shared LoRA training. (a,c,e,f) 16/8/4-bit backbones: accuracy surges to almost 100% within 600 steps, indicating that fingerprint readout is acquired. (b,d) False control with three identical 16-bit backbones (Qwen3-8B and Llama-3.1-8B only): accuracy remains at chance throughout training.

	Pointwise		Pairwise		Setwise
Model	Base	LoRA	Base	LoRA	Base	LoRA
Qwen3-8B	0.334	0.353	0.492	0.583	0.342	0.619
Llama3.1-8B-Instruct	0.341	0.378	0.487	0.589	0.381	0.612
Gemma-3-27B-IT	0.328	0.382	0.502	0.640	0.346	0.612
Qwen3-32B	0.348	0.390	0.502	0.601	0.364	0.627

Appendix

Table 15: Balanced accuracy on LoRA training to identify bit-width of the generated text.

Panel of Figure 23 — Figure 23: Confusion matrix of OOD template contexts for all models.

Panel of Figure 24 — Figure 24: Internal Awareness self-report: Abnormality, Quantization, Bitwidth for all models before and after training (shared LoRA).

Panel of Figure 25 — Figure 25: External Awareness self-report: Abnormality, Quantization, Bitwidth for all models before and after training (shared LoRA).

Explore similar work

Hallucination Is Linearly Decodable from Mid-Layer Hidden States in Quantized LLMs
May 30, 2026Aizierjiang AiersilanHidden States Large Language Model Quantization
University of Macau
Evidence for Limited Metacognition in LLMs
Date pendingChristopher AckermanMetacognition Agent Behavior Modeling
Evaluating and Improving LLM Self-Modeling
Aug 31, 2026Siqi Zeng, Andre N. Assis, Rowan WangIntrospection Adapters
University of Illinois · Constellation · Anthropic

May 30, 2026, cs.LGJ/K move · Enter open · S save

Hallucination Is Linearly Decodable from Mid-Layer Hidden States in Quantized LLMs

Aizierjiang Aiersilan

University of Macau

Hidden States Large Language Model Quantization Truth Decoding

Open paper AskPDF arXiv

We investigate whether open-source LLMs encode a linearly separable truthfulness signal in their hidden states, and at which network depth this signal is strongest. Across three

7

B--

8

B instruction-tuned models (Llama-3.1-8B, Mistral-7B, Qwen2.5-7B) loaded in

4

-bit NF4 quantization, we extract per-layer hidden states on four hallucination benchmarks (TruthfulQA, HaluEval-QA, FEVER, and a controlled synthetic set) and compare four detection approaches: linear and MLP probes, INSIDE EigenScore, self-consistency, and attention entropy. A linear probe on a single mid-network layer achieves

0.904

1.000

AUROC on held-out splits, while sampling-based detectors do not exceed

0.541

AUROC under the same protocol. The truthfulness signal is approximately linear: MLP probes rarely surpass linear probes by more than

0.01

AUROC. Peak probing layers fall in a consistent band across model families on natural-language benchmarks -- blocks~

13

18

of~

32

for Llama and Mistral, and blocks~

19

25

of~

28

for Qwen. First-block attention entropy provides a complementary signal in knowledge-grounded settings (

0.866

0.941

AUROC on HaluEval-QA) at no additional inference cost. The low discriminability of sampling methods under this protocol reflects a structural mismatch between paired-label evaluation and the information these methods access, rather than an inherent limitation of those methods. Code and data are released for full reproducibility on a single

8

,GB GPU.

Abstract

Explore similar work

Hallucination Is Linearly Decodable from Mid-Layer Hidden States in Quantized LLMs

Evidence for Limited Metacognition in LLMs

Evaluating and Improving LLM Self-Modeling