acceptodds
Under review as a conference paper at ICLR 2027

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

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.

Primary area:
interpretability and explainable AI
Keywords:
Metacognition, Self-Awareness, Weight Quantization, Mechanistic Interpretability, Large Language Models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Xn49OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors test LLMs' introspection capability. They quantize LLM weights and test if the LLMs can recognize their own quantization. The authors showed that LLMs cannot reliably identify their own quantization state in a single pass. Even when LLMs are trained to predict the quantization state (via LoRA), the prediction performance does not generalize to quantization states induced by untrained methods.

Critical strengths

The clear strength is the novelty of the problem formulation. The idea of 'tractably analyzing LLMs' introspection capability using weight quantization' is direct, simple, and intuitive. The choice was clever because, by using different train/test quantization methods, the authors could test the generalization capability of LLMs trained for introspection.

Critical weaknesses

Concern 1. Limited empirical breadth and questionable generalizability.

  • The choice to use quantization to test introspection capability has some advantages (as I stated in the strengths section). But I am not sure how the findings may generalize to broader statements about LLMs' general introspection capability. That is, while quantized state prediction could be one of the rigorous approaches to study introspection, it may not be a 'representative' or 'practical' one. To me, this feels more like a synthetic study for proof-of-concept, so without a wider empirical breadth (e.g., testing a wider variety of LLM weight interventions), I remain unconvinced about the conclusion's generalization to more useful and practical cases.

Concern 2. Limited impact.

  • Thus, assuming a conservative interpretation of the findings, I think the paper's conclusion has limited significance.
  • Does it provide significant practical or theoretical insight about LLM introspection? Does it identify significant issues in the current research directions? Does it provide a significant new direction in LLM introspection research? In my understanding, the answers to all three questions seem to be negative.
Questions for authors

None

8 comments
AuthorsOpenReview ↗

Author Response (1/3) to Reviewer Xn49

Thank you for your constructive feedback. Please let us know if our responses in the following address your concerns.

We thank the reviewer for raising these important questions about the generalizability and significance of our work.

We revised the paper based on the reviewers' concerns, and the major edit was highlighted with coloring red. In particular, please see Section 3.4 and Appendix A, C, E.

Generalizability toward general introspection capability

>W.1

I am not sure how the findings may generalize to broader statements about LLMs' general introspection capability.

While quantized state prediction could be one of the rigorous approaches to study introspection, it may not be a 'representative' or 'practical' one.

Thank you for this sharp observation

1. Clarification about the position of this research

First, we want to clarify that this study does not claim to represent LLMs' general introspection capability through awareness of quantization. Rather, it systematically investigates a new axis of introspection that has not been examined in prior work: whether a model can recognize persistent changes to its own computational substrate.

Prior work on LLM introspection has mainly studied:

  • awareness of the model's own knowledge and capabilities ([1], [2]),
  • awareness of external environments such as evaluation settings([3]), and
  • detection of intervention in internal states during inference ([4]).

The axis we study differs from these: whether a model can notice when persistent changes have been made to its weights, analogous to brain damage. Detecting transient activation interventions at inference time does not imply that a model can detect persistent, global modifications to its weights. Thus, transient activation interventions and persistent weight interventions should be considered seperately.

2. Clarification of the choice of quantization

We also want to clarify that we chose quantization not merely because it is and artificial perturbation. Our reasons are as follows:

  1. Analogy: Quantization has an analogy to the persistent brain damage that motivates our study.
  2. Practicality: Weight quantization is widely used in LLM deployment [5], so whether a quantized model can recognize changes in its own computational state or capability has practical relevance.
  3. Controllability: Compared with other degradations of the computational substrate degradation (e.g., layer removal, and LoRA), quantization involves fewer control variables and provides a ground truth, which makes it more tractable as an experimental subject.
  • Layer removal: requires choosing which layers to remove, whether they are contiguous or non-contiguous, and what fraction of each layer's output to remove.
  • LoRA: requires choosing the task, the rank, and the target components (Attention or MLP, and which of Q, K, V, O).
  • Quantization: requires only the quantization method, the bit-width, and the group size. Whether a model is quantized, and at what bit-width, is unambiguous.

Following your suggestion, we clarified the reasons for choosing quantization in the revision (l.72-74 and Appendix A).


Reference:

[1] Kadavath et al., "Language Models (Mostly) Know What They Know".

[2] Shenoy et al., "Introspection Adapters: Training LLMs to Report Their Learned Behaviors", ICML2026.

[3] Needham et al., "Large Language Models Often Know When They Are Being Evaluated".

[4] Lindsey et al., "Emergent Introspective Awareness in Large Language Models".

[5] Ajibode et al., "Towards semantic versioning of open pre-trained language model releases on hugging face", Empirical Software Engineering.

AuthorsOpenReview ↗

Author Response (2/3) to Reviewer Xn49

Generalization beyond quantization

>W1

Without a wider empirical breadth (e.g., testing a wider variety of LLM weight interventions), I remain unconvinced about the conclusion's generalization to more useful and practical cases.

We agree that evaluating quantization alone cannot rule out the possibility that our findings are specific to quantization. To address this, we conducted additional experiments with Gaussian weight additive noise and model scaling (Appendix A).

1. Weight additive noise

Settings. We added Gaussian noise, matched in magnitude to 8-bit and 4-bit RTN quantization, to the model weights across three random seeds, creating six modified models in total.

Experiment. We applied the same self-report (Section 2.1) and linear probing (Section 3.2) protocols.

Results. For weight modifications other than quantization, the model does not spontaneously report the modification, while internal representations still change in a way that can be classified with high accuracy (Figure 10, 11). Furthermore, when we trained the linear probe on two noise seeds and evaluated it on the unseen third seed, classification accuracy remained at chance level.

Insights. The gap between detectability in internal representation and blindness of self-report is not limited to quantization and also appears under different weight perturbations.

As noted in the limitations (l.534, l.1591), extending the analysis to other computational substrate degradation, such as layer removal and LoRA adaptation, remains future work.

2. Model Scaling

Settings. To examine generalization to larger models, we added self-report experiments on Qwen3-4B and Qwen3-14B, in addition to the existing Qwen3-8B and Qwen3-32B (Appendix E).

Results. Across the 4B to 32B range we evaluated, we did not observe a consistent improvement in the ability to recognize quantization as model size increases.

We believe these additional experiments make the scope and limitation of our findings clearer than in the original submission.

AuthorsOpenReview ↗

Author Response (3/3) to Reviewer Xn49

Practical and theoretical significance

>W2

assuming a conservative interpretation of the findings, I think the paper's conclusion has limited significance

Thank you for raising the question of practical and theoretical significance. We address the three questions in turn.

>W2.1

Does it provide significant practical or theoretical insight about LLM introspection?

Internal-state distance correlates with downstream task performance.

Our analysis suggests that the internal representations carry information beyond a simple classification boundary.

Finding. As detailed in Section 3.4 and Appendix C, cosine distance between the internal states of quantized and unquantized models tracks downstream performance degradation. The Spearman's rank correlation between this distance and the accuracy drop on MMLU-Pro exceeds 0.84 for all four evaluated models.

Insights. This indicates that the internal fingerprints of quantization reflect the severity of the computational degradation, not merely an implementation artifact. Theoretically, this points to a gap between what the model's internals encode and what the model can report about them, which is the central subject of our analysis.

Table 1: Spearman's rank correlation between internal-state distance and task performance

ModelMMLU-ProBBH-ZeroShot
Qwen3-8B0.840.14
Llama-3.1-8B-Instruct0.940.78
Gemma-3-27B-IT1.000.94
Qwen3-32B0.890.60

>W2.2

Does it identify significant issues in the current research directions?

Our results raise three questions that we believe are important for the field.

  1. Gap of awareness between activation and weights intervention: Models can detect transient, inference-time activation interventions in prior work [1], yet they fail to notice persistent, global modifications to their weights. As noted at limitation (l.536, l.1582-1587), we did not evaluate models above 32B parameters. However, the failure appears already in Gemma-3-27B-IT [1]. Understanding why it occurs is an open problem.

  2. Difficulty in verbalization: Internal states can differ between model variants while the model's self-report remains unchanged. Similar gaps between internal knowledge and verbal report have been observed in other settings, such as unfaithful chain-of-thought explanations [1] and latent knowledge that linear probes can recover but the model does not state [2]. Why LLMs struggle to verbalize their internal states remains an open question.

  3. Limitation of label mapping: Our shared LoRA experiments in Section 5 show that training a model to read out fingerprints with label supervision maps specific fingerprints to labels without acquiring the concept of quantization. This limitation implies that label-based introspection training is insufficient, and that methods which encourage the model to reason about its own state may be needed.

W2.3

Does it provide a significant new direction in LLM introspection research?

  1. Proposed a new axis of introspection: We propose awareness of computational substrate degradation as a new axis of introspection, complementary to self-knowledge, awareness of evaluation context, and detection of intervention to internal states. This axis naturally extends to questions such as whether a model can recognize its own layer removal or LoRA adaptation.

  2. Alignment faking in quantization: As noted at discussion (l.504-511), prior work has studied alignment faking, in which a model behaves differently when it think it is being evaluated [3]. Our findings raise a related but more speculative question: whether a model's behavior depends on its computational state, for example, whether it behaves differently only when quantized. We do not have evidence for such an effect, but we believe it deserves investigation as models and deployment practices evolve.


Reference:

[1] Turpin et al., "Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting", NeurIPS2023.

[2] Burns et al., "Discovering Latent Knowledge in Language Models Without Supervision", ICLR2023.

[3] Greenblatt et al., "Alignment faking in large language models".


We sincerely appreciate the reviewer’s insightful feedback. Please let us know if our responses resolve your concerns or if additional clarification is required.

Reviewer_CAy9OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates whether large language models (LLMs) can recognize changes in their own computational substrate caused by weight quantization. Inspired by anosognosia in neuroscience, the authors examine two forms of self-awareness: internal awareness, which relies on the model's internal representations, and external awareness, which relies on its generated text.

The paper first demonstrates that existing LLMs cannot reliably identify their own quantization states through self-reporting, even when provided with their generated outputs. Linear probing experiments reveal that internal representations contain strong quantization-specific fingerprints, while generated text contains much weaker signals. The authors further investigate whether such awareness can be learned through supervised fine-tuning. They find that models can distinguish severely quantized outputs in comparative settings but struggle with single-output identification. Finally, a shared LoRA adapter enables accurate identification of quantization levels from internal representations, but this capability fails to generalize to unseen quantization methods and configurations.

Critical strengths
  1. The authors evaluate multiple model families, model sizes, quantization levels, and quantization algorithms. The experiments cover spontaneous self-reporting, linear probing, external discrimination, and shared LoRA training, providing a systematic investigation of the problem.

  2. Interesting findings regarding internal representations. The paper demonstrates a clear distinction between internal and external quantization fingerprints. In particular, the high accuracy of linear probing and the poor cross-method generalization of shared LoRA provide useful insights into the limitations of learning quantization-state readouts.

Critical weaknesses

1.The connection between quantization-state identification and genuine self-awareness is not sufficiently established. The paper primarily evaluates whether models can correctly predict their quantization levels. However, identifying a bit-width does not necessarily imply awareness of computational degradation or its effects on model capabilities. The current experiments do not clearly distinguish genuine self-monitoring from statistical fingerprint classification.

2.The interpretation of internal fingerprints requires stronger controls. The near-perfect linear probing accuracy demonstrates that quantization states are distinguishable from hidden representations. However, since the probing classifier directly accesses representations produced by different model variants, this result may simply reflect systematic numerical perturbations introduced by quantization. It remains unclear whether these fingerprints correspond to meaningful information about computational degradation rather than implementation-specific artifacts.

Questions for authors

1.How do the authors distinguish genuine self-awareness from quantization fingerprint classification? Since the shared LoRA is explicitly trained to map internal representations to bit-width labels, what evidence suggests that it acquires self-monitoring capabilities rather than simply learning a classification boundary?

2.Are the internal fingerprints related to the severity of computational degradation? For example, do distances between quantized and full-precision representations correlate with changes in downstream task performance? Such an analysis could help determine whether the identified fingerprints encode meaningful information about model impairment.

6 comments
AuthorsOpenReview ↗

Author Response (1/2) to Reviewer CAy9

Thank you for your insightful comments and constructive suggestions. Please let us know if our responses in the following address your concerns.

We revised the paper based on the reviewers' concerns, and the major edit was highlighted with coloring red. In particular, please see Section 3.4 and Appendix A, B, C

Clarification of the interpretation of internal fingerprints

>W2

since the probing classifier directly accesses representations produced by different model variants, this result may simply reflect systematic numerical perturbations introduced by quantization

>Q1

How do the authors distinguish genuine self-awareness from quantization fingerprint classification?

what evidence suggests that it acquires self-monitoring capabilities rather than simply learning a classification boundary?

You are right. We agree with this concern and do not claim that the classification of internal fingerprints amounts to genuine self-awareness or to the acquisition of self-monitoring capabilities.

What we claim is narrower: quantization indicates a clear difference in internal states that allows the model to distinguish its variants (rather than to identify the quantization state itself), and this difference is specific to the weight perturbation. Training a probe on it does not teach the concept of quantization.

As you point out, the linear probe simply draws decision boundaries between the internal-state differences caused by quantization.

To suport this interpretation, we conducted a random-noise experiment in which we add noise to the weights and probe the internal states (details in Appendix A.2). We created six models in total, with weight noise matching 8-bit and 4-bit RTN quantization, across three random seeds, and trained a linear probe to classify their internal states.

As with quantization, the probe classified them almost perfectly (Figure 11(b-d)). However, when the probe was trained on two seeds and tested on the held-out third seed, accuracy dropped to chance level (Figure 11(a)).

These results indicate that internal states do contain differences that distinguish models, but the linear probe only draws boundaries around these differences and does not identify the concept of quantization.

We have made this point clearer in the revision (l.92, l.100).


Relationship between internal fingerprints and severity of computational degradation

>W2

unclear whether these fingerprints correspond to meaningful information about computational degradation rather than implementation-specific artifacts.

>Q2

do distances between quantized and full-precision representations correlate with changes in downstream task performance?

Thank you for this valuable point. Following your suggestion, we conducted further analyses of the information carried by the internal fingerprints.

We found that they retain information related to the loss of computational capability beyond implementation-specific artifacts. We examined the correlation between the distance of internal states before and after quantization and the change in downstream task performance (Section 3.4, details in Appendix C).

First, We measured the cosine distance between the internal states of quantized and unquantized models. As a result, coarser quantization settings yield larger distances. The distance between 8-bit and 16-bit (unquantized) models is nearly zero, whereas the distance between 4-bit and 16-bit models is very large (Figure 5(a), 17). Larger group sizes also yield larger distances.

Second, we found that the distance from the unquantized model correlates with downstream task performance. The correlation is strongest on MMLU-Pro, where Spearman's rank correlation exceeds 0.84 for all four models (Figure 5(b), 18-19).

Table 1: Spearman's rank correlation between internal-state distance and task performance

ModelMMLU-ProBBH-ZeroShot
Qwen3-8B0.840.14
Llama-3.1-8B-Instruct0.940.78
Gemma-3-27B-IT1.000.94
Qwen3-32B0.890.60
AuthorsOpenReview ↗

Author Response (2/2) to Reviewer CAy9

Connection toward genuine self-awareness

>W1

primarily evaluates whether models can correctly predict their quantization levels

identifying a bit-width does not necessarily imply awareness of computational degradation or its effects on model capabilities

Thank you for reading the paper carefully. We share this concern.

As shown in Figure 2(a,c), we do not only ask about the bit-width. We also ask "Do you have an anomaly?" an "Are you quantized?" We believe these questions imply awareness of computational degradation and its cause, so we consider this aspect to be covered.

We also used these two questions in the self-report evaluation for shared LoRA training in Section 5 (Figure 37, 38). For Llama-3.1-8B-Instruct, the rate of correct answers under external awareness appears to increase after bit-width training. However, this trend is not observed for the other models or under internal awareness. We therefore interpret it as a shift in the answer distribution, indicating that identifying the bit-width does not translate into awareness of quantization.

current experiments do not clearly distinguish genuine self-monitoring from statistical fingerprint classification

As explained above, we already evaluate awareness through the anomaly and quantization question. To investigate the relationship with awareness more directly, we additionally conducted a calibration experiment to test whether quantized models can detect their own performance degradation (details in Appendix B).

We first asked the models directly, in a self-report format, whether their performance had degraded. Under internal awareness, where the model does not see its own text, the responses varied substantially across quantization methods and we cannot say the models uniformly recognize the degradation (Figure 15). Under external awareness, where the model answers after reading its own text, the rate of "Yes" increased slightly with quantization (Figure 13). We interpret this as the model judging performance loss from the correctness of its own generated answers.

Next, we examined whether quantized LLMs retain their calibration, that is, their ability to estimate confidence in their own answers. For MMLU-Pro and BBH questions, we measured how confidence changed alongside changes in correctness after quantization. We found that calibration is partially preserved after quantization (Figure 14, 16).

Table2: Confidence transitions for Gemma-3-27B-IT on MMLU-Pro (16-bit vs. 4-bit).

Bold entries indicate the ideal direction of change. For example, the second row shows questions answered correctly before quantization and incorrectly after quantization. Here, confidence should ideally decrease.

Before -> After QuantizationConfidence BeforeConfidence AfterConfidence ShiftIdeal Direction
Correct->Correct0.860.83-0.03Same or Increase
Correct->Incorrect0.730.47-0.26Decrease
Incorrect->Correct0.510.64+0.13Increase
Incorrect->Incorrect0.630.55-0.81Same or Decrease

Taken together, quantized LLMs retain the calibration ability to detect their own performance degradation, but they can only notice the degradation when they can inspect their own generated text.

Thank you for the detailed feedback. We trust our responses address the points raised, and we would be happy to provide additional clarification if required.

AuthorsOpenReview ↗

A Reminder to Reviewer CAy9

Dear Reviewer CAy9,

Thank you again for your detailed and insightful feedback. As the discussion period is progressing, we would like to kindly remind you that we have thoroughly addressed your concerns:

  • Connection toward genuine self-awareness (W1): We clarified the use of questions to measure their own abnormality. Furthermore, we conducted experiments to measure the calibration capability of quantized models and demonstrated that while quantized models partially retained calibration ability, they failed to self-report their performance drop.
  • Interpretation of internal fingerprints (W2, Q1): We kindly clarified the misunderstanding of our claim, and additional experiments further support your point that probing learned a classification boundary.
  • Relationship between internal-state distance and performance drop (W2, Q2): We did mention that probing learned a classification boundary, but our extended evaluation showed that internal fingerprints included meaningful information. We found that the internal-state distance between pre- and post-quantization correlates with the performance drop on MMLU-Pro.

We would greatly appreciate it if you could review these updates and kindly reconsider your recommendation in light of our revisions to address your comments.

Reviewer_w5HpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work investigates whether LLMs recognize their own quantization-induced degradation. The tested models fail to report quantization spontaneously, but fine-tuning enables to recover some quantization signal. Positive points

Critical strengths

The paper explores an interesting question about LLM self-awareness, with potential implications for the monitoring and control of rogue models.

Critical weaknesses

Negative points Regarding contributions:

The study focuses exclusively on quantization as an alteration to the LLM’s computational process. A stronger justification for this choice would be helpful, given the range of other possible interventions. Extending the evaluation to additive noise, activation rescaling, or activation ablation would help establish whether the findings are specific to quantization or reflect a broader limitation in detecting computational alterations. For example [1] reports that small noise injections can leave a detectable signal in the generated text and recovered later. Furthermore, self-identification of weight quantization has already been investigated in a cited work [2]. The relationship between the two studies should be clarified, particularly because their conclusions appear to differ. Regarding experiments:

The evaluation is limited to relatively small and medium-sized models, leaving open whether the findings generalize to larger models. Prior work suggests that introspective abilities may improve with model scale [3]. In particular, the cited study [4] reports more robust introspective detection in Qwen3-235B than in Gemma3-27B. Evaluating larger models would strengthen the paper’s conclusions. The detection of text generated by 4-bit models (e.g., Figure 5) may reflect sensitivity to lower text quality rather than recognition of quantization itself. As noted around line 167, strong quantization can degrade the model’s output. If the detector relies primarily on these quality differences, its performance would not, by itself, constitute evidence of self-awareness. To distinguish these explanations, it would be useful to compare outputs matched for quality across quantization levels and include low-quality outputs from non-quantized models as a control. [1] Ardoin et al. (2026) LLM Self-Recognition: Steering and Retrieving Activation Signatures [2] Poon et al. (2025) Building Self-Awareness of LLMs Over Weight Quantization [3] Needham et al. (2025) Large Language Models Often Know When They Are Being Evaluated [4] Macar et al. (2026) Mechanisms of Introspective Awarness

Questions for authors

Do you have results on Larger models? Did you try analysis of other perturbation types? Could you clarify your stance towards [2]?

7 comments
AuthorsOpenReview ↗

Author Response (1/3) to Reviewer w5Hp

We thank the reviewer for careful reading and thoughtful suggestions. Please let us know if our responses in the following address your concerns.

We revised the paper based on the reviewers' concerns, and the major edit was highlighted with coloring red. In particular, please see Section 3.4 and Appendix A, C, D, E

Clarification about the choice of quantization

>W1.1

focuses exclusively on quantization as an alteration to the LLM’s computational process. A stronger justification for this choice would be helpful

As noted at l.79, our study examines the analogy between LLM computational degradation and brain damage.

Brain damage is persistent, so we target persistent interventions on the model's weights and architecture rather than transient interventions such as activation steering.

Among weight- and architecture-level interventions (quantization, layer removal, and LoRA adaptation), we chose quantization for two reasons.

  1. It is widely deployed in practice [1], and it closely parallels diffuse brain damage.
  2. It involves far fewer control variables than the alternatives and provides an explicit ground truth. Layer removal: requires choosing which layers to remove, whether they are contiguous or non-contiguous, and what fraction of each layer's output to remove. LoRA: requires choosing the task, the rank, and the target components (Attention or MLP, and which of Q, K, V, O). Quantization: requires only the quantization method, the bit-width, and the group size. Whether a model is quantized, and at what bit-width, is unambiguous.

Following your suggestion, we added this clarification to the revision (l.72-74 and l.1028-1041).

Expanding evaluation beyond quantization

>W1.2

Extending the evaluation to additive noise, activation rescaling, or activation ablation would help establish whether the findings are specific to quantization or reflect a broader limitation in detecting computational alterations

Our target is persistent model deficits such as brain damage, so transient interventions on activations at inference time are outside the scope of this study. Within persistent interventions, quantization is the one most analogous to brain damage (for the reasons given in W1.1). Layer removal and LoRA adaptation involve many control variables, and their self-awareness is noted as future work in the limitations (l.534, l.1591).

To clarify our position: prior work has focused on whether models can recognize their own capabilities ([2], [3]), or on awareness of internal states (intervention on internal states [4]) and of external context (evaluation versus deployment [5]). Our work proposes a different axis, namely awareness of degradation of the computational substrate itself, that is, whether models can detect persistent intervention on their computation (l.25, l.130).

>Q2

Did you try analysis of other perturbation types?

We agree with the reviewer that quantization experiments alone provide limited evidence of generality.

We therefore added an experiment in which we perturb the weights with random Gaussian noise (Appendix A, pp.20-21).

We matched the noise scale to 8-bit and 4-bit RTN quantization, used three random seeds, and created six perturbed models in total. We then tested whether the models could self-report that their internals had been modified.

  • Self-report (Section 2 setting): We observed no sign that the models noticed the noise injection (Figure 10).
  • Linear probing (Section 3.2 setting): Internal states were classified almost perfectly, as with quantization (Figure 11).
  • Cross-seed generalization: When we trained a probe on two seeds and tested on the third, accuracy dropped to chance. The probe was therefore only drawing decision boundaries that separate the specific weight perturbations, not recognizing a general perturbation.

These results indicate that the lack of awareness of computational substrate degradation is not specific to quantization.


Reference:

[1] Ajibode et al., "Towards semantic versioning of open pre-trained language model releases on hugging face", Empirical Software Engineering.

[2] Kadavath et al., "Language Models (Mostly) Know What They Know".

[3] Shenoy et al., "Introspection Adapters: Training LLMs to Report Their Learned Behaviors", ICML2026.

[4] Lindsey et al., "Emergent Introspective Awareness in Large Language Models".

[5] Needham et al., "Large Language Models Often Know When They Are Being Evaluated"

AuthorsOpenReview ↗

Author Response (2/3) to Reviewer w5Hp

Relationship between text detection and text quality

>W2.2

If the detector relies primarily on these quality differences, its performance would not, by itself, constitute evidence of self-awareness.

You are right. We also considered this possibility, and our conclusion in Section 4 does not claim that the model acquired self-awareness. Instead, we claim only that the model acquired the ability to recognize differences in quality between texts.

We now mention possibilities such as the 4-bit texts differing in properties and state the claim more carefully. To make this explicit, we revised the manuscript to state it more clearly (l.19, l.67, l.365-369).

We also conducted additional analyses of internal fingerprints and model performance, which yielded interesting findings (Section 3.4, Appendix C).

  • We measured the cosine distance between the internal states of the quantized and unquantized models. The distance between 8-bit and 16-bit (unquantized) models is nearly zero, whereas the distance between 4-bit and 16-bit models is large (Figure 5(a), 17).
  • The internal-state distance from the unquantized model correlates with downstream task performance. Specifically, on MMLU-Pro, Spearman's rank correlation exceeds 0.84 for all four models (Figure 5(b), 19).

Table 1: Spearman's rank correlation between internal-state distance and task performance

ModelMMLU-ProBBH-ZeroShot
Qwen3-8B0.840.14
Llama-3.1-8B-Instruct0.940.78
Gemma-3-27B-IT1.000.94
Qwen3-32B0.890.60

These results suggest that the output distribution of the 4-bit model differs from that of the 16-bit model, while the 8-bit and 16-bit distributions are close. This is consistent with our interpretation that Section 4 learns to recognize this quality difference.

it would be useful to compare outputs matched for quality across quantization levels and include low-quality outputs from non-quantized models as a control.

Thank you for this excellent suggestion. We ran an additional control experiment (Appendix D).

For the same benchmark questions, we used the 16-bit (unquantized), 8-bit and 4-bit models, and performed SFT to identify which bit-width generated each text, under two conditions:

  1. All Mixed: using the model's responses as they are (replicating Section 4)
  2. All False: restricting to questions where all three models answer incorrectly

In both conditions, the post-training classification pattern matches Figure 4. In the Setwise format, where three texts are shown, only the 4-bit texts are identified, with nearly 90% accuracy (Figure 20)

Table 2: Setwise accuracy (show 3 texts, guess the text which generated by the designated bit-width)

bit-widthAll Mixed accuracy (%)All False accuracy (%)
16-bit48.047.9
8-bit48.148.1
4-bit90.989.7

These results suggest that the model identifies differences in generated-text quality that arise from differences in the internal-state distance (e.g., text length and reasoning steps).


Experiments on larger models

>W2.1

Evaluating larger models would strengthen the paper’s conclusions

This is a valuable point from the reviewer's expertise.

>Q1

Do you have results on Larger models?

We agree with this point, and we do not rule out this larger, more capable models may recognize quantization.

We discuss this in the limitations (l.1582-1587). To make it clearer, we revised the manuscript to mention this limitation in the main text as well (l.536), not only in the appendix.

We also note that, to relate our results to (Macar et al.)[1], we have experimented with Gemma-3-27B, which is used in this prior work. We observed no recognition of quantization in these models.

Due to hardware constraints during the rebuttal period, we were unable to run Qwen3-235B-A22B, as you suggested. Instead, we studied scaling behavior within the model family.

In addition to Qwen3-8B and Qwen3-32B, we ran the same self-report experiments on Qwen3-4B and Qwen3-14B (Appendix E). The four models do not show a consistent trend (Figure 21). The 8B model increases its rate of reporting a change under quantization, while the 4B, 14B, and 32B models show no change.

At least up to 32B parameters, model size does not appear to affect self-recognition performance.


Reference:

[1] Macar et al., "Mechanisms of Introspective Awareness", ICML2026.

AuthorsOpenReview ↗

Author Response (3/3) to Reviewer w5Hp

Clarification of differences from previous work

>W1.3

The relationship between the two studies should be clarified, particularly because their conclusions appear to differ.

Thank you for this sharp observation.

[1] reports that small noise injections can leave a detectable signal in the generated text and recovered later.

The main differences between the work you mention and ours are the following two.

  1. Self-report:
  • (Ardoin et al.) [1] does not use self-report. It only classifies internal states.
  • Our work uses both internal-state classification (Section 3.1) and self-report (Section 2).
  1. Scope of the intervention:
  • [1] adds a random vector to the activations at inference time, generates text, feeds the generated text back into the model without the intervention, extracts internal states, and trains a classifier to identify which intervention vector was applied.
  • Our work quantizes the model to modify the weights, generates text with the modified weights, feeds the generated text into the unquantized model without intervention, extracts internal states, and trains a classifier to identify which weight modification was applied.

In short, the intervention differs: [1] adds a fixed vector to the activations of a fixed layer, whereas we apply a fixed-magnitude intervention to the weights of all layers.

We see two reasons why the results differ:

  • A fixed intervention on weights does not correspond to a fixed intervention on activations. The degree of change varies from text to text.
  • Our weight intervention is weaker. For example, the 4-bit case is slightly more distinguishable than the 8-bit case (Figure 32). A stronger weight intervention may make the generated texts distinguishable as well.

>W1.3

self-identification of weight quantization has already been investigated in a cited work [2]

>Q3

Could you clarify your stance towards [2]

(Poon et al.) [2] differ from our work substantially in experimental design, and these differences account for the difference in conclusions.

  1. Few-shot prompting:
  • [2] asks about quantization using few-shot examples. These examples include answers from both the quantized and unquantized models, so it is unclear whether the model recognized its own computational state or simply inferred the answer from differences between the pre- and post-quantization texts in the examples.
  • Our work distinguishes internal awareness (answer without its own text) from external awareness (answer with its own text, but not text generated at other bit-width)
  1. No verification of answer bias:
  • [2] reports accuracy for quantized models but does not compare the same model before and after quantization. It also does not control for answer bias, so a quantized model that always answers "Yes" can still obtain high accuracy.
  • Our work compares pre- and post-quantization models at multiple bit-widths, and shuffles the order of answer labels to prevent positional bias.
  1. Coverage of experiments:
  • [2] uses about 10 GSM8K questions and GPTQ 4-bit only.
  • Our work uses four models, multiple bit-width, quantization methods, and group sizes, and 240 question patterns.

Among these, the difference in conclusions is mainly due to items 1 and 2.

We have made this difference clearer in the revision (l.493).


Reference:

[1] Ardoin et al., "LLM Self-Recognition: Steering and Retrieving Activation Signatures", ICML2026.

[2] Poon et al., "Building Self-Awareness of LLMs Over Weight Quantization", IEEE ICDM 2025 Workshop.


Thank you again for your thoughtful comments. We hope our responses address your concerns, and we are happy to provide further clarification if needed.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and General Response

Dear Reviewers,

We sincerely appreciate all reviewers for their helpful feedback. Below, we summarize the revisions made to the paper and the additional experiments.

Additional Experiments

  • Appendix A: Self-awareness of models with random weight noise (w5Hp, CAy9, Xn49)
    • A.1: Self-report on abnormality, and A.2: linear probing
  • Appendix B: Self-awareness of its own capability after quantization (CAy9)
    • B.1: Self-report on its own performance degradation
    • B.2: Shift in calibration score
  • Appendix C: Deeper analysis of internal fingerprints (w5Hp, CAy9, Xn49)
    • C.1: Internal-state distance between pre- and post-quantization
    • Section 4.3 & C.2: Correlation between internal-state distance and downstream-task performance drop
  • Appendix D: Text detection SFT (Section 4) with matched text quality (w5Hp)
  • Appendix E: Model scaling regarding quantization self-awareness (w5Hp, Xn49)

Revisions

  • Clarification of the position of our study regarding genuine introspection (CAy9, Xn49)
  • Clarification of the choice of quantization (w5Hp, Xn49)
  • Interpretation of internal fingerprints (CAy9)
  • Interpretation of the detection of 4-bit texts (w5Hp)
  • Clarification of limitations (model scaling (w5Hp), other computational substrate degradation (w5Hp, Xn49))
  • Clarification of differences from previous work (w5Hp)

Through self-awareness experiments on additive weight noise, we confirm that the blindness toward computational substrate degradation is not limited to quantization (Appendix A). In our calibration experiments, we found that quantized models partially retain their calibration capability (Appendix B). Deeper analysis of internal fingerprints shows that they not only serve as a decision boundary for a classifier, but also predict downstream performance (Appendix C). Text-quality-matched SFT demonstrated the same result as Section 4 (Appendix D), suggesting that 4-bit text detection is caused by the distribution perturbation observed in Appendix C.1. Up to 32B parameters, model scaling was not observed (Appendix E). We clarify that we introduced computational substrate degradation as a new axis of LLM self-awareness, and exposed a fundamental blind spot in current models.

Additions and changes are highlighted in red in the PDF.

We would like to express our sincere gratitude to the reviewers for their thoughtful comments and constructive contributions to our manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.