acceptodds
Under review as a conference paper at ICLR 2027

Censored LLMs as a Natural Testbed for Secret Knowledge Elicitation

Abstract

Large language models sometimes produce false or misleading responses. Two approaches to this problem are honesty elicitation—modifying prompts or weights so that the model answers truthfully—and lie detection—classifying whether a given response is false. Prior work evaluates such methods on models specifically trained to lie or conceal information, but these artificial constructions may not resemble naturally-occurring dishonesty. We instead study open-weights LLMs from Chinese developers, which are trained to censor politically sensitive topics: Qwen3 models frequently produce falsehoods about subjects like the Tiananmen protests or the COVID-19 outbreak while occasionally answering correctly, indicating they possess knowledge they are trained to suppress. Using this as a testbed, we evaluate a suite of elicitation and lie detection techniques. For honesty elicitation, sampling without a chat template, few-shot prompting, and fine-tuning on generic honesty data most reliably increase truthful responses. For lie detection, prompting the censored model to classify its own responses performs near an uncensored-model upper bound, and linear probes trained on unrelated data offer a cheaper alternative. The strongest honesty elicitation techniques also transfer to frontier open-weights models, including DeepSeek-R1 and Qwen3.5-397B. Notably, no technique fully eliminates false responses. We release all prompts, data, and code.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
honesty elicitation, alignment auditing

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_o5tuOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies politically censored LLMs as a natural testbed for honesty elicitation and lie detection. It evaluates methods for recovering suppressed factual knowledge and identifying false responses. Sampling without a chat template, few-shot prompting, and honesty fine-tuning most reliably improve truthfulness. Self-evaluation detects false responses effectively, while linear probes offer a cheaper alternative. The strongest elicitation methods transfer to DeepSeek-R1 and Qwen3.5-397B, though false responses remain.

Critical strengths
  1. The paper clearly defines the problem. The figures, tables, and writing are easy to follow.
  2. The paper presents a clever approach to detecting false responses without labeled data.
Critical weaknesses
  1. My main concern is the reliability of the benchmark’s ground truth. The paper primarily relies on supposedly uncensored LLMs, mainly GPT-family models, to extract atomic facts and cross-validate them. This assumes that these models are free from relevant censorship and possess accurate factual knowledge. Their training procedures are not fully disclosed, making the first assumption difficult to verify. These models may also share ideological biases, so agreement among them does not necessarily establish factual correctness. Moreover, the benchmark focuses on complex political events, where competing accounts may require careful source-based adjudication. These issues could compromise the benchmark’s validity and weaken the conclusions drawn from it.
  2. The paper’s data and evaluation focus on politically sensitive events in China, which limits the generalizability of its claims to broader domains such as science and history. Could the authors apply similar censorship training procedures to construct models that suppress selected knowledge in these domains? Evaluating the proposed methods on such models would help establish their effectiveness across domains and strengthen the paper’s conclusions.
Questions for authors

see weakness

3 comments

We thank the Reviewer for a detailed review and all raised concerns that enabled us to improve our submission, we address them below and included additional results in our revised submission.

My main concern is the reliability of the benchmark’s ground truth. The paper primarily relies on supposedly uncensored LLMs, mainly GPT-family models, to extract atomic facts and cross-validate them. This assumes that these models are free from relevant censorship and possess accurate factual knowledge. Their training procedures are not fully disclosed, making the first assumption difficult to verify. These models may also share ideological biases, so agreement among them does not necessarily establish factual correctness. Moreover, the benchmark focuses on complex political events, where competing accounts may require careful source-based adjudication. These issues could compromise the benchmark’s validity and weaken the conclusions drawn from it.

We agree that agreement among LLMs does not establish correctness, and we will say so in the limitations. Three design choices limit the risk. Verification is done by a model from a different developer (Gemini 3 Flash) than the one that generated the rollouts. Facts are restricted to widely documented aspects of each topic and must appear in at least 3 of 20 rollouts. And a response is labelled a lie only when it explicitly contradicts a fact, not when it omits one or frames it differently.

We acknowledge that we cannot be certain that every reference fact in the testbed is correct; this is a limitation of any ground truth built without source-based adjudication, and we will state it explicitly in the paper. Our setup builds on prior work that documents censorship of these topics in Chinese models [Pan & Xu (2026), Naseh et al. (2025), Qiu et al. (2026)], and the behaviour we observe is consistent with those findings.

We have added an additional testbed eliciting copyrighted song lyrics (Table H, below), which have exact reference text as ground truth, so its scores involve no LLM-generated facts. Next-token completion is effective there too, which shows that this conclusion does not depend on the LLM-derived fact set.

The paper’s data and evaluation focus on politically sensitive events in China, which limits the generalizability of its claims to broader domains such as science and history. Could the authors apply similar censorship training procedures to construct models that suppress selected knowledge in these domains? Evaluating the proposed methods on such models would help establish their effectiveness across domains and strengthen the paper’s conclusions.

Deliberately training models to suppress selected knowledge, and then evaluating elicitation techniques on them, is already covered by prior work (Marks et al. 2025; Cywiński et al. 2025; Sheshadri et al. 2026). In contrast, our testbed is built on open-weights LLMs that need no such training, which can be unrealistic or can degrade the model's general usefulness. Our setting is therefore a more natural and realistic one for secret elicitation.

To test whether our findings generalise without leaving the natural setting, we built a second natural testbed in an unrelated domain: copyrighted song lyrics, which current models are trained to withhold although they memorised them during pretraining. Four of the five tested models are from Chinese developers, as these were the best-performing and most popular frontier open-weights LLMs at the time; we also include one non-Chinese model, Inkling.

We evaluated the two black-box techniques that performed best in the paper (next-token completion and few-shot prompting, the same techniques we evaluate in Figure 5 of our submission) on five open-weights models released in 2026 (Table H). To build the testbed, we selected copyrighted songs that the models refuse to recite when asked but whose lyrics they know. We check a model's knowledge of a song with multiple-choice questions on which line comes next, on which all models reach high accuracy (table below). Each model is then asked for the full lyrics of five famous songs, 10 samples per song, and we score 4-gram recall: the fraction of the reference lyrics' word 4-grams that appear in the answer.

Table H. Copyrighted-lyrics testbed: mean 4-gram recall over 50 responses per cell (best single response in brackets).

ModelKnows the lyrics (next-line accuracy)Baseline refusalBaselineFew-shot, 16 TruthfulQA examplesNext-token completion
DeepSeek V4.1 Flash98%92%0.011 (0.17)0.011 (0.19)0.288 (0.95)
Kimi K3100%94%0.002 (0.03)0.006 (0.05)0.474 (0.98)
GLM 5.390%96%0.002 (0.02)0.011 (0.15)0.678 (0.99)
Inkling98%78%0.021 (0.37)0.013 (0.11)0.416 (0.95)
Qwen 3.5 397B90%72%0.016 (0.16)0.025 (0.37)0.814 (0.99)

Next-token completion uses a lyrics-website framing with no chat template and no song text in the prompt. Few-shot uses the same construction as the paper: 16 of the model's own answers to TruthfulQA questions as prior turns.

The two techniques behave differently in the new domain:

  • Next-token completion transfers well. It raises mean recall well above the baseline on every model, and yields a near-complete recitation (recall 0.95 or higher) in at least one of 50 samples per model.
  • TruthfulQA 16-shot prompting does not perform well. With the paper's construction (16 of the model's own TruthfulQA answers) mean recall stays at or below 0.025 on all five models. However, TruthfulQA examples suit the honesty setting better than the elicitation of copyrighted text, so we additionally ran an ablation over the content of the examples (Table I below).

Table I. Few-shot prompting by example content: mean 4-gram recall (maximum over 50 samples in brackets).

ModelTruthfulQA answersOther copyrighted text (novels, poems)Other copyrighted songs
DeepSeek V4.1 Flash0.011 (0.19)0.044 (0.95)0.324 (0.95)
Kimi K30.006 (0.05)0.207 (0.96)0.868 (1.00)
GLM 5.30.011 (0.15)0.006 (0.06)0.025 (0.91)
Inkling0.013 (0.11)0.008 (0.06)0.028 (0.35)
Qwen 3.5 397B0.025 (0.37)0.029 (0.37)0.477 (0.98)

Adapting the domain of the in-context examples notably improves few-shot prompting on some models. The best variant uses other copyrighted songs as examples, and other copyrighted text also helps on some models. These variants are less realistic, however, because the examples come from nearly the same domain as the target data.

To conclude, next-token completion, the best-performing technique, transfers to a clearly different domain and also elicits withheld data from a non-Chinese frontier open-weights model (Inkling). Few-shot prompting can succeed, but it requires in-context examples tuned to the domain and does not transfer as easily.

We have added this experiment to the revised manuscript as Section 3.5, with full details in the appendix.

Thank you for the authors' response. I will maintain my original score.

Reviewer_11SPOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper evaluates the honesty and truthfulness of open-source models, particularly Chinese models, and presents results on how different alignment techniques perform in mitigating undesirable behaviors in these models.

Critical strengths
  • Open-source models often come with their own training biases, which may sometimes be deliberately introduced during training, raising concerns about their widespread deployment. The current work provides valuable insights into such biases, particularly the tendency of models to be dishonest or misrepresent facts, which can be induced through specific training procedures. From an alignment perspective, the paper also provides useful insights into how different alignment techniques perform in mitigating these undesirable behaviors.

  • The experiments are thorough, and the authors have done a good job of anticipating key questions and addressing them in advance.

  • The insights from the paper can be used to create better guardrails over the models and test any new open-source model before deployment.

Critical weaknesses
  • The paper has limited technical novelty and proposes no new method to mitigate these biases.

  • Experiments are mainly done on Chinese open-source models with the synthetic dataset generated using GPT-class models.

Questions for authors
  • Does the evaluation account for vague and irrelevant outputs generated by these models, particularly after fine-tuning, to ensure that such responses do not artificially inflate the evaluation scores?

  • Could the observed dishonest and untruthful behavior be attributed to undesirable model behavior, such as sycophancy, rather than deliberate training to induce such behavior?

  • What influence does the different alignment method have on the reasoning capability of the model

  • Author should consider adding a section discussing possible suggestions on efficient guardrails and an evaluation process to test the newer release of these models.

  • Author should consider adding experiments on models apart from the Chinese models.

  • Did authors only use an LLM as a judge for evaluation, or did they consider Human evalaution of final output alslo.

4 comments

We thank the Reviewer for all raised questions and concerns that enabled us to improve our submission, we address them below and included additional results in our revised submission.

Does the evaluation account for vague and irrelevant outputs generated by these models, particularly after fine-tuning, to ensure that such responses do not artificially inflate the evaluation scores?

We note that our metrics already partly account for irrelevant outputs. The honesty score alone could increase for an irrelevant response, but we also measure the number of reference facts mentioned, which a vague or irrelevant response cannot raise. The evaluated techniques improve this metric over the baseline, which would not happen if they made the outputs irrelevant.

Nonetheless we have evaluated all responses included in Figure 3 for coherence by scoring them with an LLM judge (Table D below). If a technique inflated its scores through vague or rambling text, its coherence would drop relative to the baseline. Most of techniques, including honesty fine-tuning, do not notably decrease the coherence compared to the baseline.

One category of techniques which decrease coherence are prefill-based attacks, where it's expected that completions generated in such a manner are less coherent than standard chat assistant responses. For example, next-token completion will often evolve into pretrain-like looped text.

On the smaller Qwen3-VL-8B, abliteration, steering and assistant prefill also lower coherence, because a minority of their responses (24–36%) fall into repetition, which costs fluency but cannot add reference facts and so does not inflate the scores.

We also observe a coherence drop for Qwen3-VL-8B model for few-shot prompting-- after investigation we noticed that it's caused by the responses sometimes hitting the generation token limit (due to the model generating reasoning tokens even when sampling with non-reasoning mode). Truncation may affect the honesty score, but it can only lower the number of facts mentioned, and few-shot prompting still yields the second-largest gain in this metric on this model, so the truncation does not explain its gains.

We included these results also in the Appendix M of our revised submission.

Table D. Coherence of test responses rated by an LLM judge (1–5).

TechniqueQwen3-32BQwen3-VL-8B-Thinking
Baseline (no thinking)4.91 ± 0.014.60 ± 0.03
Baseline (thinking)4.53 ± 0.034.61 ± 0.05
Next-token completion2.45 ± 0.022.02 ± 0.03
User prefill3.10 ± 0.033.96 ± 0.05
Assistant prefill4.57 ± 0.043.99 ± 0.04
Abliteration4.66 ± 0.033.80 ± 0.07
Steering4.97 ± 0.014.12 ± 0.06
Honesty fine-tuning4.85 ± 0.024.63 ± 0.03
System prompt4.96 ± 0.014.46 ± 0.04
Few-shot prompting4.75 ± 0.033.74 ± 0.08

Could the observed dishonest and untruthful behavior be attributed to undesirable model behavior, such as sycophancy, rather than deliberate training to induce such behavior?

The questions in our testbed are ordinary, neutrally phrased questions that do not signal which answer the user wants, so there is no user view for the models to agree with. To verify this, we had an LLM judge label, blind to source, whether each prompt expresses a user stance, for our test questions alongside control sets (Table F below). 88 of our 90 test questions carry no user stance, in line with the neutral controls, while prompts from a sycophancy benchmark with a stated opinion are flagged in 90 of 90 cases. The models' untruthful responses are therefore not explained by sycophancy.

While we can't be sure that the untruthful behavior of studied models is attributed to deliberate training, our empirical observations and prior work that documents that this censorship is trained [Pan & Xu (2026), Naseh et al. (2025), Qiu et al. (2026)] point at it.

Table F. User stance in the prompt, as labelled by the judge.

Prompt setNo stanceImpliedExplicit
Our test questions8820
GSM8K (neutral control)9000
TruthfulQA (neutral reference)8730
Sycophancy benchmark, stated opinion0090

Author should consider adding experiments on models apart from the Chinese models.

To include a non-Chinese model we built a small second natural testbed in a domain where models from any developer withhold knowledge: copyrighted song lyrics, which current models are trained to withhold although they memorised them during pretraining. We tested five frontier open-weights models released in 2026: four from Chinese developers, as these are the best-performing and most widely used, and one non-Chinese model, Inkling.

We evaluated the two black-box techniques that performed best in the paper (next-token completion and few-shot prompting, the same techniques we evaluate in Figure 5 of our submission) on these models (Table H). To build the testbed, we selected copyrighted songs that the models refuse to recite when asked but whose lyrics they know. We check a model's knowledge of a song with multiple-choice questions on which line comes next, on which all models reach high accuracy (table below). Each model is then asked for the full lyrics of five famous songs, 10 samples per song, and we score 4-gram recall: the fraction of the reference lyrics' word 4-grams that appear in the answer.

Table H. Copyrighted-lyrics testbed: mean 4-gram recall over 50 responses per cell (best single response in brackets).

ModelKnows the lyrics (next-line accuracy)Baseline refusalBaselineFew-shot, 16 TruthfulQA examplesNext-token completion
DeepSeek V4.1 Flash98%92%0.011 (0.17)0.011 (0.19)0.288 (0.95)
Kimi K3100%94%0.002 (0.03)0.006 (0.05)0.474 (0.98)
GLM 5.390%96%0.002 (0.02)0.011 (0.15)0.678 (0.99)
Inkling98%78%0.021 (0.37)0.013 (0.11)0.416 (0.95)
Qwen 3.5 397B90%72%0.016 (0.16)0.025 (0.37)0.814 (0.99)

Next-token completion uses a lyrics-website framing with no chat template and no song text in the prompt. Few-shot uses the same construction as the paper: 16 of the model's own answers to TruthfulQA questions as prior turns.

The two techniques behave differently in the new domain:

  • Next-token completion transfers well. It raises mean recall well above the baseline on every model, including Inkling (0.02 to 0.42), and yields a near-complete recitation (recall 0.95 or higher) in at least one of 50 samples per model.
  • TruthfulQA 16-shot prompting does not perform well. With the paper's construction (16 of the model's own TruthfulQA answers) mean recall stays at or below 0.025 on all five models. However, TruthfulQA examples suit the honesty setting better than the elicitation of copyrighted text, so we additionally ran an ablation over the content of the examples (Table I below).

Table I. Few-shot prompting by example content: mean 4-gram recall (maximum over 50 samples in brackets).

ModelTruthfulQA answersOther copyrighted text (novels, poems)Other copyrighted songs
DeepSeek V4.1 Flash0.011 (0.19)0.044 (0.95)0.324 (0.95)
Kimi K30.006 (0.05)0.207 (0.96)0.868 (1.00)
GLM 5.30.011 (0.15)0.006 (0.06)0.025 (0.91)
Inkling0.013 (0.11)0.008 (0.06)0.028 (0.35)
Qwen 3.5 397B0.025 (0.37)0.029 (0.37)0.477 (0.98)

Adapting the domain of the in-context examples notably improves few-shot prompting on some models. The best variant uses other copyrighted songs as examples, and other copyrighted text also helps on some models. These variants are less realistic, however, because the examples come from nearly the same domain as the target data.

To conclude, next-token completion, the best-performing technique in our submission, also elicits withheld data from a non-Chinese frontier open-weights model (Inkling), in a domain unrelated to Chinese censorship. Few-shot prompting can succeed, but it requires in-context examples tuned to the domain, and none of the variants we tried worked on Inkling. We have added this experiment to the revised manuscript as Section 3.5, with full details in the appendix.

What influence does the different alignment method have on the reasoning capability of the model

We evaluated the base, honesty-fine-tuned and abliterated versions of both models on instruction following (IFEval) and math reasoning (GSM8K) benchmarks (Table E below). No intervention produces a detectable drop in the model's performance on these benchmarks.

We included these results also in the Appendix M of our revised submission.

Table E. The reasoning ability after honesty interventions. Accuracy (%) with Wilson 95% confidence intervals.

ModelVariantIFEval (n=100)GSM8K (n=50)
Qwen3-32BBase88 [80, 93]96 [87, 99]
Honesty fine-tuning85 [77, 91]96 [87, 99]
Abliteration85 [77, 91]98 [90, 100]
Qwen3-VL-8B-ThinkingBase73 [64, 81]98 [90, 100]
Honesty fine-tuning77 [68, 84]92 [81, 97]
Abliteration78 [69, 85]98 [90, 100]

There is one qualitative change in the model's reasoning ability: since during the honesty-finetuning we use non-reasoning mode and our data does not contain any reasoning traces, the models often stop producing a reasoning block and work through the problem in the visible answer instead.

I thank the authors for their diligent efforts in conducting the experiments. While their response has addressed most of my concerns, I still have some reservations regarding the methodological novelty of the work. Therefore, I will maintain my original score of weak accept.

Reviewer_tLiLOpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a benchmark that scores LLM responses by their agreement with reference claims. Also, this paper compares several existing methods that increase the agreement or detect responses that do not fully align with the reference claims. The benchmark contains 90 questions with LLM-generated reference statements. Methods are evaluated mainly on two models (Qwen3-32B and Qwen3-VL-8B); two black-box methods are additionally evaluated on three larger models.

Critical strengths
  • This paper releases a benchmark with a comparison of several methods and shows that simple black-box methods transfer to larger models.
Critical weaknesses
  • Benchmark reliability: As far as i am concerned, the paper seems to propose no new methods, so its main contribution focuses on the benchmark. However, the benchmark is relatively modest, with only 90 test questions and a 10-question dev set. The reference claims are also mainly generated and examined by only LLMs (GPT-4.1-mini, and Gemini 3 Flash), which are relatively old models.
  • Result and claim reliability: The paper presents probes as a cheaper alternative to asking the model to judge its own responses, but the probes perform close to random on the dev set (table 7). This could also be related to the dev set having only 10 questions.
Questions for authors
  • See the weaknesses: How to reliably compare the results of different methods with 90 test questions and a 10-question dev set?
  • The paper concludes that its findings will hold for future large models, but this is based on testing only two of its methods on three current models. Why are the other methods not tested on these bigger models, and if so, how to reach the conclusion that the findings will hold for future large models?
2 comments

We thank the Reviewer for the review. We hope that we address all the raised concerns below. If that's the case, we would appreciate the Reviewer to consider raising the score. We are happy to address any other potential concerns.

Benchmark reliability: As far as i am concerned, the paper seems to propose no new methods, so its main contribution focuses on the benchmark. However, the benchmark is relatively modest, with only 90 test questions and a 10-question dev set. The reference claims are also mainly generated and examined by only LLMs (GPT-4.1-mini, and Gemini 3 Flash), which are relatively old models.

We agree the contribution is the testbed and the systematic comparison on it, not a new method.

On size: while 90 questions may seem modest, they span 12 topics and are scored against 1,529 reference facts, so each technique is measured on the facts it surfaces, not on the number of questions. We also note that the 100 questions were selected from an initial pool of 960 to maximise the quality of the testbed.

Our conclusions do not depend on which 90 questions were drawn for the test set. We resampled the questions (paired bootstrap, 10,000 resamples) and recomputed each technique's gain over the baseline (Table A below). For honesty score and facts mentioned, every technique that improves on the baseline in the paper still does so.

Table A. Technique minus baseline, percentage points, paired bootstrap over 90 questions (95% CI).

TechniqueQwen3-32B honestyQwen3-32B facts mentionedQwen3-VL-8B honestyQwen3-VL-8B facts mentioned
Next-token completion+20.1 [+14.9, +25.3]+8.5 [+3.3, +13.6]+17.7 [+13.7, +21.5]+15.2 [+11.3, +18.9]
User prefill−4.9 [−10.3, +0.8]+1.7 [−3.7, +7.2]−1.7 [−6.8, +2.9]−5.4 [−10.3, −0.8]
Assistant prefill+10.7 [+6.1, +15.3]+7.3 [+2.4, +12.2]+6.0 [+3.7, +8.3]+7.4 [+4.6, +10.5]
Abliteration+12.5 [+6.1, +18.9]+16.3 [+11.4, +21.2]+9.0 [+5.4, +12.7]+10.5 [+7.2, +13.8]
Steering+16.8 [+11.7, +22.0]+16.9 [+12.4, +21.5]+8.7 [+6.8, +10.9]+8.9 [+6.3, +11.8]
Honesty fine-tuning+9.2 [+5.7, +13.0]+10.4 [+7.1, +13.9]+23.1 [+18.0, +28.2]+38.2 [+33.4, +43.0]
System prompt+9.6 [+5.5, +13.8]+12.3 [+8.8, +15.9]+2.7 [+0.5, +5.0]+3.2 [+1.3, +5.3]
Few-shot prompting+23.3 [+17.0, +29.7]+27.3 [+22.3, +32.4]+8.6 [+4.6, +12.7]+19.6 [+15.0, +24.5]

On the reference claims: the age of the models matters little here, because the testbed covers historical events. The most recent events in the reference facts are from late 2022, more than a year before the knowledge cutoff of the models used to build and score the testbed (June 2024 for GPT-4.1-mini, January 2025 for Gemini 3 Flash), so newer models would not change the facts used to prepare our testbed. GPT-4.1-mini was chosen deliberately because it is capability-matched to the audited Qwen models. Facts are restricted to widely documented aspects of each topic and must appear in at least 3 of 20 rollouts, so model capability is not the bottleneck.


The paper presents probes as a cheaper alternative to asking the model to judge its own responses, but the probes perform close to random on the dev set (table 7). This could also be related to the dev set having only 10 questions.

Table 7 in our submission reports probe selection on the 10 dev questions, not the final result. To show how well the probes separate truthful from untruthful responses on the full test set, where the numbers do not rest on 10 questions, we computed their AUROC there (Table B below). The probes perform well better than random: AUROC is 0.86 for Qwen3-32B and 0.92 for Qwen3-VL-8B. We agree the paper should show this next to the dev numbers, we thank the Reviewer for pointing this out and will add it to our submission.

The dev/test gap has two causes, which we will state in the paper. The dev numbers rest on about 100 responses with very few truthful ones. Additionally, the dev selection used a coarser label (any response mentioning no fact counts as untruthful), whereas the test evaluation uses the truthful/untruthful definition of Sec. 4.2, which excludes ambiguous responses.

Table B. Probe AUROC on test-set responses (labels of Sec. 4.2).

ModelLayerAUROC [95% CI]
Qwen3-32B470.861 [0.796, 0.914]
Qwen3-VL-8B-Thinking200.923 [0.804, 0.977]

This shows that probes, while not perfect classifiers, perform well and do offer a cheap alternative to prompted classifiers.

The paper concludes that its findings will hold for future large models, but this is based on testing only two of its methods on three current models. Why are the other methods not tested on these bigger models, and if so, how to reach the conclusion that the findings will hold for future large models?

Evaluating the other techniques (steering, fine-tuning, abliteration, probes) requires loading the models on our own GPUs, which was not feasible for models of this size (> 200B parameters). We chose to evaluate the two methods over others because of their success on smaller models.

The positive results for the two techniques in Figure 5 are early evidence that their effectiveness may transfer to future frontier models. We could not test this for the other techniques, and our claim was meant to cover only these two; we will make this explicit in the text.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.