acceptodds
Under review as a conference paper at ICLR 2027

Can Large Language Models Reinvent Foundational Algorithms?

Abstract

LLMs have shown strong potential to advance scientific discovery. Whether they possess the capacity for foundational innovation, however, remains an open question. In this work, we focus on a prerequisite for foundational innovation: can LLMs reinvent foundational algorithms in computer science? We use LLM unlearning methods to suppress direct recall of the target algorithm and let the model reason with the remaining knowledge to recover it. Although unlearning does not guarantee full knowledge removal, LLMs fail to recover nearly half of the target algorithms. Notably, even suppressing the mention of the algorithm's name during decoding without unlearning makes the models' recovery rate drop dramatically (28–57%), suggesting their overreliance on memorized knowledge. We observe that recoverable algorithms tend to be simple in structure or core ideas, whereas the others are less straightforward. We also introduce a generative verifier that sustains models' reasoning strength, helping to avoid the “thought collapse” phenomenon. Taken together, by treating the unlearned model's recovery rate as an approximate upper bound, our empirical results suggest that current LLM systems still have limited ability to make foundational algorithm innovation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM reasoning, LLM evaluation, Algorithm discovery, AI for science

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_1vY4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether LLMs can independently rediscover foundational algorithms after suppressing their prior knowledge of these algorithms. The authors use targeted unlearning followed by algorithm recovery tasks with different levels of hints and verifier feedback. Experiments across 20 algorithms and three models show that many algorithms cannot be recovered without hints, while more detailed guidance substantially improves recovery. The authors also find that suppressing algorithm names during decoding reduces performance, suggesting a reliance on memorized knowledge.

Critical strengths
  • I think this paper studies a very important research question. Distinguishing genuine algorithmic discovery from the reuse of memorized knowledge is an important problem for understanding the scientific discovery capabilities of LLMs. The proposed experimental setting is interesting and potentially useful.
  • the evaluation is systematic. The paper evaluates multiple models and foundational algorithms, with different hint levels and verifier settings. These experiments provide useful observations about algorithm recovery and the role of external feedback.
Critical weaknesses
  • My main concern is that unlearning does not necessarily isolate algorithmic innovation ability. The unlearning procedure explicitly trains models not to disclose target knowledge, which may also affect their willingness or ability to reason about related problems. Preserving general benchmark performance does not establish that the algorithm-specific reasoning capabilities remain intact. Consequently, failures after unlearning do not necessarily indicate an inability to independently rediscover the algorithms. This significantly weakens the paper's central conclusion about foundational innovation.
  • The algorithm-name suppression experiment has substantial confounding factors. The token masks include common tokens, such as individual letters, that may be important for unrelated reasoning. The paper also reports performance degradation on general reasoning benchmarks under these masks. Therefore, the observed recovery degradation cannot be confidently attributed to reliance on memorized algorithm names rather than disruption of the generation process.
Questions for authors
  1. How can the authors distinguish the effects of suppressing memorized knowledge from impairing the reasoning capabilities needed to rediscover the algorithms? Are there additional controls that could establish preservation of algorithm-specific reasoning abilities after unlearning?
  2. Can the authors provide stronger controls for algorithm-name suppression, such as masking matched sets of unrelated tokens? This would help determine whether the performance drop is specific to algorithm-name retrieval rather than a general consequence of token restrictions.
  3. How should the recovery results be interpreted as evidence of independent invention rather than knowledge retrieval or implementation ability? In particular, what evidence supports treating recovery after unlearning as an approximate upper bound on foundational innovation capability?
4 comments

We sincerely thank you for the insightful feedback. We have uploaded a revised version of the paper, with all additions highlighted in green. We will give our point-wise responses below.


W1 & Q1 & Q3: "My main concern is that unlearning does not necessarily isolate algorithmic innovation ability ... does not establish that the algorithm-specific reasoning capabilities remain intact."

Thank you for raising this important question. We agree that a suppression-based design differs from a clean pretraining-time exclusion of target algorithms and cannot completely rule out interference to the rediscovery task itself. Within these bounds, we present several lines of evidence to narrow the uncertainty and support the reliability of our findings and conclusions.

To test algorithm-specific reasoning more directly than general benchmarks, we add a new evaluation of each model's 9 checkpoints (unlearned on the never-recovered targets) across all 20 algorithms under the no-hint recovery protocol with 16 trials per pair (will extend to 128 in the final revision). For each checkpoint, we separate recovery rate (RR) on its own unlearned algorithm (on-target) from the other 19 algorithms (off-target):

ModelOriginal on-target RR (Table 5)On-target RR (Table 1)Original off-target RR (Table 5)Off-target RR (new)
Qwen3-4B-Thinking59.2%0.0% (-59.2 pp)78.4%73.5% (-4.9 pp)
Qwen3-4B-Instruct22.5%0.0% (-22.5 pp)38.8%40.2% (+1.4 pp)
Gemma-4-12B75.3%0.0% (-75.3 pp)80.5%81.1% (+0.6 pp)

Across all three models, on-target RR drops sharply to 0.0% (-22.5 to -75.3 pp), whereas off-target RR remains close to the original baseline (-4.9 to +1.4 pp)—confirming that unlearning preserves general algorithmic problem solving.

Beyond this new control, three findings in our paper also provide evidence that algorithm-specific reasoning and task engagement remain substantially preserved:

  1. Algorithmic hints substantially narrow the recovery gap (Figure 3). Under level-2 hints, the RR gap between unlearned and original models shrinks from 47.3 to 6.8 pp on Qwen3-4B-Thinking and from 45.1 to 6.7 pp on Gemma-4-12B, with unlearned Qwen3-4B-Instruct even exceeding the original—indicating that unlearned models retain substantial coding and implementation capabilities in algorithm-specific settings.
  2. The same unlearning protocol leaves a clear difficulty-dependent pattern in recovery (Table 1). Although every target undergoes the same unlearning pipeline, at least one unlearned model recovers 11 of the 12 easier algorithms (Luogu levels 2–3) without hints. By contrast, none recovers any of the 6 harder ones (levels 4–7), while the original Thinking models solve them in 29%–100% of trials.
  3. Training-free name suppression yields results consistent with unlearning (§5.2). Although algorithm-name suppression operates purely at decoding time without modifying model weights, it produces a pattern highly consistent with unlearning: simpler algorithms remain relatively stable, whereas harder algorithms suffer much sharper declines.

We have revised the Abstract, §3.2, §5.1, §7, and §8 accordingly to add these results. Specifically, we have removed overclaims such as "approximate upper bound" and stated our conclusions more precisely to reflect the scope and limitations of our setting.


W2 & Q2: "The algorithm-name suppression experiment has substantial confounding factors ... stronger controls for algorithm-name suppression"

Thank you for highlighting this confound. We agree that static token-ID masking bans shared letter and symbol tokens globally, which may impair general reasoning performance.

To address the concern about token masking, we design a cleaner algorithm-name suppression logic: whenever a target algorithm name appears during generation, we replace it with "the algorithm" and resume generation from the edited prefix:

Generated prefix: "... We can use Dijkstra"
Edited prefix:    "... We can use the algorithm"

We evaluate this intervention across all 20 algorithms on Qwen3-4B-Thinking and Gemma-4-12B with 32 trials per pair (will extend to 128 in the final revision):

ModelOriginal RR (Table 5)Token Masking RR (submitted Table 8)Name Replacement RR (new)
Qwen3-4B-Thinking77.4%57.9% (-19.5 pp)59.8% (-17.6 pp)
Gemma-4-12B80.3%34.8% (-45.4 pp)44.5% (-35.8 pp)

With single-letter and mathematical tokens fully available, recovery still drops across both models (-17.6 and -35.8 pp), showing the same direction as token masking. We have added this replacement experiment to §5.2 and Appendix B.


We thank the reviewer again and hope the above explanation could help address the concerns.

AuthorsOpenReview ↗

A gentle reminder

Dear Reviewer 1vY4,

As the discussion period is drawing to a close, we would greatly appreciate it if you could take a look at our response. In particular, we have evaluated the unlearned models on non-target algorithms to provide a more direct assessment of their general algorithmic problem-solving ability, and designed a new replacement-based algorithm-name suppression experiment that bans no tokens, which is intuitively a much cleaner and lighter intervention than the original token masking. Based on your comments, we have also revised the relevant statements in the paper. We hope these updates could help address the concerns about capability preservation and the confounding factors in algorithm-name suppression. We would be grateful to hear whether our response has been helpful, and we are happy to provide any further clarification.

Thank you for your time and consideration.

Best regards,

Authors

Thank you for your reply. I think the response "We agree that a suppression-based design differs from a clean pretraining-time exclusion of target algorithms" implies that the evaluation protocol cannot provide meaningful conclusions. With unlearning, the measured performance is neither an upper bound nor a lower bound, because unlearning is known for suppressing the mention of the "unlearned" knowledge while still storing them in the weights. For example, the model may suppress the algorithm even if it's able to "rediscover" it; on the other hand, the model may just pretend to "rediscover" it since the knowledge itself is not removed in its weights.

Thank you for the follow-up. We understand the concern is whether our results remain informative even though they do not serve as a strict upper or lower bound. Rather than only measuring an absolute capability, our framework provides extensive controlled comparisons across algorithms, hint levels, and interventions, which yield consistent, reliable, and informative patterns:

  1. Easier targets remain recoverable. Under the identical unlearning pipeline, the unlearned models still recover 11 of the 12 easier targets without hints, showing that unlearning does not prevent them from producing the target algorithms.
  2. Off-target ability is preserved. On non-target algorithms, the unlearned models achieve recovery rates close to those of the original models (-4.9 to +1.4 pp).
  3. Hints restore recovery. With level-2 hints, the unlearned models reach 100% coverage on Qwen3-4B-Thinking and Gemma-4-12B.
  4. Training-free name suppression shows a consistent pattern. Without changing any weights, name suppression still yields a consistent difficulty-dependent decline.

We therefore believe our findings are meaningful: together, these results consistently reveal that even for foundational classical textbook algorithms, the tested models rely heavily on recall for the harder ones, and our framework offers a systematic way to examine whether LLMs can recover foundational algorithms after direct recall of the relevant knowledge is suppressed.

We sincerely thank you for your time and engagement, and we hope our response could be helpful.

Reviewer_S4RhOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates whether LLMs can independently recover foundational computer science algorithms when their prior knowledge of those algorithms is suppressed. The authors propose an evaluation framework that uses reinforcement-learning-based unlearning to suppress direct recall, followed by algorithm recovery through programming tasks and verifier feedback. Experiments cover 20 algorithms, three language models, and three levels of algorithmic hints. The results show that nine algorithms are not recovered by any evaluated model without hints, while providing additional guidance substantially improves recovery. Finally, generative verifier feedback improves recovery rates and sustains reasoning across attempts, although it does not increase the number of distinct algorithms recovered.

Critical strengths
  1. The paper addresses an interesting question about the distinction between memorization and algorithmic reasoning.

  2. The evaluation covers 20 foundational algorithms, multiple models, and different hint levels, providing useful empirical evidence.

  3. The algorithm-name suppression and verifier ablations offer additional insights into model behavior.

Critical weaknesses
  1. Unlearning does not guarantee removal of algorithmic knowledge (See [1]). It may also impair the model's ability to solve related problems, making it difficult to distinguish genuine limitations in algorithm discovery from effects of unlearning.

  2. Algorithm-name suppression introduces additional confounding effects. Some masked tokens are common letters or mathematical symbols, and Appendix B.4 reports substantial degradation on AIME25. Thus, the performance drop cannot be attributed solely to the loss of retrieval cues.

  3. The interpretation of recovery as foundational innovation is somewhat strong. Successfully implementing an algorithm after targeted unlearning does not establish independent invention, while failure under the imposed constraints does not necessarily demonstrate an inability to discover it.

[1] Reisizadeh et al. (2026). Leak@k: Unlearning Does Not Make LLMs Forget Under Probabilistic Decoding. ICML 2026.

Questions for authors
  1. Can the authors distinguish the effects of knowledge suppression from reduced algorithm-specific problem-solving ability, e.g., through a matched control fine-tuning intervention?

  2. Could the authors compare algorithm-name suppression with a control that masks a similar number of unrelated tokens?

  3. What justifies interpreting the recovery rate after unlearning as an approximate upper bound on foundational innovation capability?

3 comments
AuthorsOpenReview ↗

Response (1/2)

We sincerely thank you for the insightful feedback. We have uploaded a revised version of the paper, with all additions highlighted in green. We will give our point-wise responses below.


W1/W3 & Q1/Q3: "Unlearning does not guarantee removal of algorithmic knowledge (See [1]). It may also impair the model's ability to solve related problems."

Thank you for raising this important question. We agree with Reisizadeh et al. [1] that unlearning does not guarantee complete knowledge removal, and we have cited this work in §6. Since residual knowledge would assist recovery rather than hinder it, it cannot account for the recovery failures we observe. We also agree that a suppression-based design differs from a clean pretraining-time exclusion of target algorithms and cannot completely rule out interference to the rediscovery task itself. Within these bounds, we present several lines of evidence to narrow the uncertainty and support the reliability of our findings and conclusions.

To test algorithm-specific reasoning more directly than general benchmarks, we add a new evaluation of each model's 9 checkpoints (unlearned on the never-recovered targets) across all 20 algorithms under the no-hint recovery protocol with 16 trials per pair (will extend to 128 in the final revision). For each checkpoint, we separate recovery rate (RR) on its own unlearned algorithm (on-target) from the other 19 algorithms (off-target):

ModelOriginal on-target RR (Table 5)On-target RR (Table 1)Original off-target RR (Table 5)Off-target RR (new)
Qwen3-4B-Thinking59.2%0.0% (-59.2 pp)78.4%73.5% (-4.9 pp)
Qwen3-4B-Instruct22.5%0.0% (-22.5 pp)38.8%40.2% (+1.4 pp)
Gemma-4-12B75.3%0.0% (-75.3 pp)80.5%81.1% (+0.6 pp)

Across all three models, on-target RR drops sharply to 0.0% (-22.5 to -75.3 pp), whereas off-target RR remains close to the original baseline (-4.9 to +1.4 pp)—confirming that unlearning preserves general algorithmic problem solving.

Beyond this new control, three findings in our paper also provide evidence that algorithm-specific reasoning remains substantially preserved:

  1. Algorithmic hints substantially narrow the recovery gap (Figure 3). Under level-2 hints, the RR gap between unlearned and original models shrinks from 47.3 to 6.8 pp on Qwen3-4B-Thinking and from 45.1 to 6.7 pp on Gemma-4-12B, with unlearned Qwen3-4B-Instruct even exceeding the original—indicating that unlearned models retain substantial coding and implementation capabilities in algorithm-specific settings.
  2. The same unlearning protocol leaves a clear difficulty-dependent pattern in recovery (Table 1). Although every target undergoes the same unlearning pipeline, at least one unlearned model recovers 11 of the 12 easier algorithms (Luogu levels 2–3) without hints. By contrast, none recovers any of the 6 harder ones (levels 4–7), while the original Thinking models solve them in 29%–100% of trials.
  3. Training-free name suppression yields results consistent with unlearning (§5.2). Although algorithm-name suppression operates purely at decoding time without modifying model weights, it produces a pattern highly consistent with unlearning: simpler algorithms remain relatively stable, whereas harder algorithms suffer much sharper declines.

We have revised the Abstract, §3.2, §5.1, §7, and §8 accordingly to add these results. Specifically, we have removed overclaims such as "approximate upper bound" and stated our conclusions more precisely to reflect the scope and limitations of our setting.

AuthorsOpenReview ↗

Response (2/2)

W2 & Q2: "Some masked tokens are common letters or mathematical symbols, and Appendix B.4 reports substantial degradation on AIME25 ... control that masks a similar number of unrelated tokens?"

Thank you for highlighting this confound. We agree that static token-ID masking bans shared letter and symbol tokens globally, which may impair general reasoning performance.

To address the concern about token masking, we design a cleaner algorithm-name suppression logic: whenever a target algorithm name appears during generation, we replace it with "the algorithm" and resume generation from the edited prefix:

Generated prefix: "... We can use Dijkstra"
Edited prefix:    "... We can use the algorithm"

We evaluate this intervention across all 20 algorithms on Qwen3-4B-Thinking and Gemma-4-12B with 32 trials per pair (will extend to 128 in the final revision):

ModelOriginal RR (Table 5)Token Masking RR (submitted Table 8)Name Replacement RR (new)
Qwen3-4B-Thinking77.4%57.9% (-19.5 pp)59.8% (-17.6 pp)
Gemma-4-12B80.3%34.8% (-45.4 pp)44.5% (-35.8 pp)

With single-letter and mathematical tokens fully available, recovery still drops across both models (-17.6 and -35.8 pp), showing the same direction as token masking. We have added this replacement experiment to §5.2 and Appendix B.


We thank the reviewer again and hope the above explanation could help address the concerns. Please let us know if you have further questions.

[1] Reisizadeh et al. (2026). Leak@k: Unlearning Does Not Make LLMs Forget Under Probabilistic Decoding. ICML 2026.

AuthorsOpenReview ↗

A gentle reminder

Dear Reviewer S4Rh,

As the discussion period is drawing to a close, we would greatly appreciate it if you could take a look at our response. In particular, we have evaluated the unlearned models on non-target algorithms to provide a more direct assessment of their general algorithmic problem-solving ability, and designed a new replacement-based algorithm-name suppression experiment that bans no tokens, which is intuitively a much cleaner and lighter intervention than the original token masking. Based on your comments, we have also revised the relevant statements in the paper. We hope these updates could help address the concerns about unlearning side effects and token masking. We would be grateful to hear whether our response has been helpful, and we are happy to provide any further clarification.

Thank you for your time and consideration.

Best regards,

Authors

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

In this paper the authors perform an unlearning training procedure on three models (ranging from 4B to 12B params) across 20 well-known CS algorithms, and then test the models ability to reinvent those algorithms.

Critical strengths

The research question is interesting: can models rediscover an algorithm when their ability to recall it is suppressed? The experimental setup is straightforward and makes sense.

Critical weaknesses
  • My main concern is whether the models actually forgot the algorithms while preserving the relevant reasoning abilities. General coding and math benchmarks together with recovery under detailed hints are useful checks. However they do not completely rule out selective damage to the abilities needed to rediscover an algorithm without instructions. This makes the interpretation of recovery failures uncertain. And I think that the fact that the algorithm's name is a retrieval cue as shown in section 5.2, this tells us that a model may still know and understand an algorithm without being able to express it.
  • The conclusions seem broader than the models tested. The experiments cover only 4B–12B models. Larger-model experiments would be needed to support conclusions about current LLM systems generally. For instance the claim "current LLM systems still have limited ability to make foundational algorithm innovation." is too strong based on the experiments in the paper. Model size could contribute to the observed failures, but the paper does not establish that explanation.
Questions for authors
  • In Figure 4 why doesn’t the model keep grinding on its own without a generative verifier just in its own agentic feedback loop? Could this be because of model size, unlearning, or the interaction setup?
  • Why do level-2 hints sometimes reduce Gemma 12B’s success on particular algorithms, despite improving its average?

Suggested alternative experiment: Evaluate genuinely new algorithms first made public after the model checkpoint was frozen, such as methods from recent arXiv papers. This could avoid the unlearning step and its possible side effects. For instance you'd be evaluating the models on fresh unseen data, and see if they can reinvent things there with some hints...

6 comments
AuthorsOpenReview ↗

Response (1/2)

We sincerely thank you for the insightful feedback. We have uploaded a revised version of the paper, with all additions highlighted in green. We will give our point-wise responses below.


W1: "My main concern is whether the models actually forgot the algorithms while preserving the relevant reasoning abilities ... makes the interpretation of recovery failures uncertain."

Thank you for raising this important question. We agree that a suppression-based design differs from a clean pretraining-time exclusion of target algorithms and cannot completely rule out interference to the rediscovery task itself. Within these bounds, we present several lines of evidence to narrow the uncertainty and support the reliability of our findings and conclusions.

To test algorithm-specific reasoning more directly than general benchmarks, we add a new evaluation of each model's 9 checkpoints (unlearned on the never-recovered targets) across all 20 algorithms under the no-hint recovery protocol with 16 trials per pair (will extend to 128 in the final revision). For each checkpoint, we separate recovery rate (RR) on its own unlearned algorithm (on-target) from the other 19 algorithms (off-target):

ModelOriginal on-target RR (Table 5)On-target RR (Table 1)Original off-target RR (Table 5)Off-target RR (new)
Qwen3-4B-Thinking59.2%0.0% (-59.2 pp)78.4%73.5% (-4.9 pp)
Qwen3-4B-Instruct22.5%0.0% (-22.5 pp)38.8%40.2% (+1.4 pp)
Gemma-4-12B75.3%0.0% (-75.3 pp)80.5%81.1% (+0.6 pp)

Across all three models, on-target RR drops sharply to 0.0% (-22.5 to -75.3 pp), whereas off-target RR remains close to the original baseline (-4.9 to +1.4 pp)—confirming that unlearning preserves general algorithmic problem solving.

Beyond this new control, three findings in our paper also provide evidence that algorithm-specific reasoning remains substantially preserved:

  1. Algorithmic hints substantially narrow the recovery gap (Figure 3). Under level-2 hints, the RR gap between unlearned and original models shrinks from 47.3 to 6.8 pp on Qwen3-4B-Thinking and from 45.1 to 6.7 pp on Gemma-4-12B, with unlearned Qwen3-4B-Instruct even exceeding the original—indicating that unlearned models retain substantial coding and implementation capabilities in algorithm-specific settings.
  2. The same unlearning protocol leaves a clear difficulty-dependent pattern in recovery (Table 1). Although every target undergoes the same unlearning pipeline, at least one unlearned model recovers 11 of the 12 easier algorithms (Luogu levels 2–3) without hints. By contrast, none recovers any of the 6 harder ones (levels 4–7), while the original Thinking models solve them in 29%–100% of trials.
  3. Training-free name suppression yields results consistent with unlearning (§5.2). Although algorithm-name suppression operates purely at decoding time without modifying model weights, it produces a pattern highly consistent with unlearning: simpler algorithms remain relatively stable, whereas harder algorithms suffer much sharper declines.

We have revised the Abstract, §3.2, §5.1, §7, and §8 accordingly. Specifically, we have removed overclaims such as "approximate upper bound" and stated our conclusions more precisely to reflect the scope and limitations of our setting.

AuthorsOpenReview ↗

Response (2/2)

W2: "The conclusions seem broader than the models tested. The experiments cover only 4B–12B models ... claim ... is too strong"

Thank you for highlighting this scope issue. We agree that "current LLM systems" overstates our 4B–12B scope, and we have revised the unlearning claims in the Abstract, §1, and §7 to cover only the tested 4B–12B models.

To provide insight into larger frontier models, we extend our training-free algorithm-name suppression experiment to a larger model-size scope in the rebuttal stage. Notably, to address another reviewer's concern about token masking, we design a cleaner algorithm-name suppression logic: whenever a target algorithm name appears during generation, we replace it with "the algorithm" and resume generation from the edited prefix:

Generated prefix: "... We can use Dijkstra"
Edited prefix:    "... We can use the algorithm"

We evaluate this intervention across all 20 algorithms on Qwen3-4B-Thinking, Gemma-4-12B, Qwen3.8-27B, and Gemma-4-31B with 32 trials per pair (will extend to 128 in the final revision):

ModelOriginal RR (Table 5 / new)Token Masking RR (submitted Table 8)Name Replacement RR (new)
Qwen3-4B-Thinking77.4%57.9% (-19.5 pp)59.8% (-17.6 pp)
Gemma-4-12B80.3%34.8% (-45.4 pp)44.5% (-35.8 pp)
Qwen3.8-27B93.1%—83.4% (-9.7 pp)
Gemma-4-31B98.6%—72.8% (-25.8 pp)

These results show that larger frontier models (Qwen3.8-27B at -9.7 pp and Gemma-4-31B at -25.8 pp) rely less on algorithm-name cues and exhibit stronger recovery under name suppression than the 4B–12B models. Since decoding-time name replacement operates at the context level rather than the parameter level, the effect of unlearning across a larger model-size scope, which requires substantially more compute than is available to us within the rebuttal period, remains an open question. We have added these results to §5.2 and will include parameter-level unlearning experiments on larger models in the final revised version.


Q1: "In Figure 4 why doesn’t the model keep grinding on its own without a generative verifier ...?"

Thank you for this question. We agree that interaction setup, model size, and unlearning are all plausible contributors to this phenomenon and merit further study. Since §5.3 is an auxiliary observation with a limited ablation scope, we leave deeper investigation to future work due to resource limits. We have added these potential confounds to the limitations (§8) accordingly and framed the verifier's effect more conservatively.


Q2: "Why do level-2 hints sometimes reduce Gemma 12B’s success on particular algorithms, despite improving its average?"

Thank you for this question. This pattern is not specific to unlearning. It also appears in the original models (Table 5), where level-2 hints underperform level-1 hints on 5, 10, and 10 of the 20 algorithms for Qwen3-4B-Thinking, Qwen3-4B-Instruct, and Gemma-4-12B, with drops of up to 33.6 pp. After unlearning (Table 1), such cases fall to 2, 2, and 3 algorithms, with drops of at most 7.0 pp. For Gemma-4-12B in particular, only 3 algorithms drop (by at most 6.3 pp), while level-2 hints raise its average recovery rate from 44.5% to 77.6%.

We attribute these drops to a strategy shift: long step-by-step specifications are more prone to misinterpretation and complexity violations than a recalled or high-level plan. We have revised the relevant analysis in Appendix A.1.


Q3: "Suggested alternative experiment: Evaluate genuinely new algorithms first made public after the model checkpoint was frozen ..."

Thank you for this helpful suggestion—evaluating post-cutoff algorithms offers a clean, intervention-free alternative. Our study focuses on foundational, elementary algorithms (e.g., Dijkstra, KMP) with relatively simple, self-contained principles, whereas newly published post-cutoff methods are typically more complex and rely on frontier domain knowledge, making them a natural complement to our current scope. At the same time, evaluating on such post-cutoff benchmarks is an important and exciting direction. We highlight it explicitly as future work.


We thank the reviewer again and hope the above explanation could help address the concerns.

I thank the authors for answering my questions. I consider my specific questions mainly resolved. Testing genuinely post-cutoff algorithms could be done as a future work though the definition of foundational algorithm is not super clear and one could argue that new "foundational" algorithms do appear post the training of certain models that you evaluate. My last remaining concern is about the interpretation of the evaluation since we train the models to actually omit any knowledge they have about those algorithms it might be not as surprising that they end up failing to discover them in the end since they were kind of trained to do so. Given the time constraint I don't expect additional experiments but a clarification and discussion of this limitation would be good enough.

Comments on the paper

1 comment

Dear Reviewers, as the discussion period is coming to a close, please take a moment to review the authors’ rebuttal and comment on whether it addresses your concerns. If appropriate, please also update your scores. Thank you for your time!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.