acceptodds
Under review as a conference paper at ICLR 2027

Anti-General Intelligence: Specialization via Focal Forgetting and Hyperplasticity

Abstract

On many ML tasks, frontier performance is achieved by pretrained LLMs that are subsequently specialized to a task by post-training on data from that task. Specialization is typically assumed to come at the cost of forgetting previous abilities, which is welcome in a specialist deployed for a single purpose. However, recent work has shown that post-trained models may retain information even without explicit anti-forgetting measures. Across a wide range of domains, we observe that small post-trained LLMs do indeed retain their pre-specialization performance on non-target tasks, most completely when specialized with RL. We analyze the characteristics of non-forgetting in a saturated specialist, finding that task performance is carried by sparse task-selective subnetworks and that non-target tasks hold capacity the specialist never reclaims. We then study an intervention for the focal forgetting of task information by *Low-Rank Ascent*, finding that forgetting alone does not reliably help the target on its own, and develop *AntiGen*, a method that improves models beyond saturation by selectively increasing model plasticity in the non-target representation subspaces that ascent identifies. Applied to small specialists, *AntiGen* outperforms direct specialization and contemporary plasticity baselines on math and commonsense reasoning tasks, generalizes better in-domain, and focuses the loss of log-probabilities in ascent-derived subspaces.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Specialization, LLM, Continual Learning, Forgetting, Catastrophic Forgetting, Post-training, Plasticity, Task Adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dvNqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors observe that small specialist LLMs trained with RL retain much of their performance on non-target tasks. They propose Low-Rank Ascent (LRA) to induce focal forgetting, and ANTIGEN, which uses the subspaces identified by ascent to host high-plasticity parameters that continue improving the target past GRPO's plateau. Results are reported on Countdown, DeepScaleR, and OpenBookQA.

Critical strengths
  • The phenomenon is interesting. RL specialists retaining non-target capabilities runs against the common catastrophic-forgetting narrative.
  • The analysis is broad: MLP activation maps, Fisher ratios, gradient cosines, parameter Jaccard overlap, and CKA.
Critical weaknesses
  • The central mechanism is not supported by the authors' own ablations. The claim is that ANTIGEN gains come from recruiting non-target subspaces identified by ascent. But plain LoRA (60.8 ± 2.8) and random addresses (59.5 ± 2.2) are close to ANTIGEN (62.4 ± 1.1), with heavily overlapping error bars. Appendix C.3 shows random addresses and plain LoRA also compress late-layer donor Fisher ratios (by 27% and 18%, versus 16% for ANTIGEN). The authors themselves write that ascent-derived addresses are not required for this effect. This contradicts the subspace-recruitment narrative.

  • Peak-based reporting on the test split. Figure 3 and Table 6 report the maximum over checkpoints on the 500-item evaluation split. Selecting the best checkpoint on the test set inflates methods with noisy trajectories. Checkpoint selection should use a separate validation split, and final or window-mean results should be reported. Comparison windows also vary across arms (1620–1800, 2000, 2400), and arms are shown only where they cover the full window, which adds another selection layer.

Questions for authors
  • Is the gap between ANTIGEN and plain LoRA or random addresses statistically significant with more seeds or a paired test?
  • Can checkpoints be selected on a held-out validation split, and what are the final-checkpoint results?
2 comments
AuthorsOpenReview ↗

Response to Reviewer dvNq

We thank the reviewer for their careful review of our work. Their review is succinct but identifies our strengths and gaps with great attention to detail. We address these gaps point by point, with new results collected at https://anonymous.4open.science/r/antigen-rebuttal-31EB/, and we hope the reviewer will consider recommending acceptance.

Improvement over our ablations

  • Our ablations are novel in the context of RL specialization, and our claims on AntiGen's performance and mechanism are not contradicted by them. Concretely, our ablations establish the following.
  • The low-rank intervention improves over GRPO at every saturated specialist, the paper's main result, and AntiGen does so most reliably (Sections R1–R4).
  • With frozen ascent-derived addresses, AntiGen is cheaper to train than high-LR LoRA, which performs comparably. In the Countdown setting, AntiGen with the frozen A matrix updates 67.2M parameters (8.9% of the model), whereas plain LoRA at the same rank updates 113.6M (15.1%). Thus, AntiGen is cheaper than this ablation but reliably achieves a similar or slightly better accuracy.
  • Consistency. Against frozen random addresses, AntiGen ends higher on Countdown (57.2 ± 2.9 against 53.4 ± 3.9, five seeds) and is the only variant whose end-of-training gain over continued GRPO is significant (p = 0.023; Q1).
  • Focal forgetting. Donor log-probability loss is 0.23 nats with the ascent-derived addresses against 0.09 with random ones, and with random addresses donors lose less than unrelated tasks (as in the submitted Fig. 4: −0.147 on donors against −0.097 on probes for AntiGen, −0.074 against −0.084 for random addresses; Section R7).
  • Mechanism. Our ablations separate two effects: the gain over continued GRPO comes from the high-learning-rate container, while the ascent-derived addresses determine what that gain costs. With AntiGen's frozen ascent-derived A, the container trains 41% fewer parameters than LoRA and directs forgetting toward the donor tasks, as above. On the late-layer Fisher ratios, Appendix C.3 compares arms at different steps: AntiGen's 16% compression is measured at step 2000, the 27% and 18% for LoRA and random addresses at step 2400, and compression grows with training time. At the matched step 2000, LoRA compresses 8%, half of AntiGen's 16%. Compression as such is a property of every high-plasticity container, as the appendix says; the address-specific effect is its size at a matched step and the log-probability profile.

Q1: Significance with more seeds

  • We ran five intervention seeds per container arm from the same specialist (Section R4). AntiGen has the highest end-of-training accuracy (final checkpoint 57.2 ± 2.9, against 56.2 ± 3.5 for LoRA and 53.4 ± 3.9 for random addresses), the smallest seed spread, and is the only variant whose end-of-training gain over continued GRPO is significant: +5.9 at the final checkpoint (Welch p = 0.023) and +6.6 over the last five checkpoints (p = 0.010), against +4.9 (p = 0.055) for LoRA and +2.1 (n.s.) for random addresses.
  • The direct comparisons between variants are not significant at this sample size: AntiGen − LoRA is +1.0 at the final checkpoint (p = 0.62; paired on seed, p = 0.47) and AntiGen − random addresses is +3.8 (p = 0.13; paired, p = 0.21). We therefore claim the gain over continued GRPO (also +2.0 ± 0.3 and +4.8 ± 0.4 on DeepScaleR at 0.8B and 2B, paired p < 0.01; Section R2), the end-of-training reliability, the parameter saving and the forgetting profile, but not a significant accuracy margin over the ablations.

Q2: Peaks, windows, and validation

  • We agree that peak test accuracy alone is a fragile statistic. The revision reports the mean of the last five checkpoints and the final checkpoint of every method over the same training steps, and moves peaks to the appendix; methods of the former Table 6 that stop earlier are compared with continued GRPO over the same steps.
  • We also show the full test learning curves in Sections R1–R4, as we do for DeepScaleR and OpenBookQA in the main text, which makes checkpoint selection moot. On Countdown, AntiGen's mean accuracy is above that of continued GRPO at nearly every checkpoint after the first 200 steps (Section R4). Since continued GRPO is flat, even its best test checkpoint, an upper bound on any validation-based selection, lies below AntiGen's final checkpoint.
  • Validation-based selection gives the same picture. On a held-out Countdown split (a disjoint generation seed), validation and test accuracy are strongly correlated across checkpoints, and the validation-selected test accuracy of each container, averaged over seeds, is above the best checkpoint of every continued-GRPO run. On OpenBookQA, the dataset's validation split reproduces the test result over three seeds (Section R3). AntiGen itself never selects on an evaluation split: its schedule is driven by the training reward.

I have read authors' response and kept my score.

Reviewer_9yEsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors ask whether small specialist LLMs waste capacity on unrelated tasks and whether inducing forgetting on them helps target-task RL. After testing Low-Rank Ascent (LRA) on donor tasks, they propose ANTIGEN: freeze LRA input matrices A, train output matrix B at a higher learning rate with GRPO, and periodically merge B into the weights.

Critical strengths
  • The large Countdown ablation (Tab. 6, Tab. 7) is useful for practitioners because testing dozens of forgetting variants and RL plasticity baselines shows that explicit forgetting alone does not help target RL.
  • Countdown gains in Fig. 3b are interesting: ANTIGEN (along with its plain-LoRA and random-A variants in the same high-learning-rate adapter setup) clearly outperforms continued GRPO and recent RL plasticity baselines such as ProRL, Clip-Cov, and ReDo.
Critical weaknesses
  • Contradiction between claims and experimental results. Reading Appendix B honestly feels like reading a rebuttal to the main paper: the authors never merge LRA adapters in the main runs (do_merge is false), report that merging before training hurts every suite tried, and show A barely changing from random initialization (well below their warning line in Fig. 15a). Plain LoRA and frozen random A also get nearly the same window mean as ANTIGEN (Tab. 6), and in Appendix C they admit ascent-derived addresses are unnecessary. They even write in Appendix C that math tasks rise slightly in every adapter arm, while mentioning the math ratio as dropping in three of four arms in Fig. 24.
  • Limited technical novelty. Once LRA initialization proves unnecessary, ANTIGEN just combines three existing LoRA ideas: a larger learning rate on B than A from LoRA+ [1], periodic merge-and-reset from ReLoRA [2], and frozen random input matrices A from asymmetric LoRA [3].
  • Missing related works and comparison baselines. The paper missed related works such as [1-3] and on learning versus forgetting [4] showing that tuning only self-attention or MLP gate and up projections learns new skills with minimal forgetting. On DeepScaleR and OpenBookQA (Fig. 5, Fig. 6), they omit plain LoRA and random A, and start from the base model instead of a saturated specialist (only one specialist-start seed appears in Fig. 16).
  • Data leakage in the AI2-RC zero-shot transfer claim. On OpenBookQA (Section 4.2, Fig. 6), the authors claim zero-shot transfer to AI2-RC, yet state in the same sentence that they selected the donor suite using LRA performance on AI2-RC itself. Worse, Appendix D and Fig. 14 show that AI2-RC (MC) (the multiple-choice version of the same ARC-Challenge benchmark) is one of the five OpenBookQA donor tasks used to build A, while being omitted from the Donors group in Fig. 6.

[1]. LoRA+: Efficient Low Rank Adaptation of Large Models. Soufiane Hayou, et. al. ICML 2024.

[2]. ReLoRA: High-Rank Training Through Low-Rank Updates. Vladislav Lialin, et. al. ICLR 2024.

[3]. Asymmetry in Low-Rank Adapters of Foundation Models. Jiacheng Zhu, et. al. ICML 2024.

[4]. How To Teach Large Multimodal Models New Skills. Zhen Zhu, et. al. ECCV 2026.

Questions for authors
  1. Starting from saturated specialists on DeepScaleR and OpenBookQA, how does ANTIGEN compare over three seeds against plain LoRA and frozen random A?
  2. Does the AI2-RC gain hold if AI2-RC (MC) is removed from the OpenBookQA donors and donors are chosen without evaluating on AI2-RC?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 9yEs

We thank the reviewer for their meticulous review of our work. The reviewer accurately summarizes our method and raises important points of clarification, particularly about our ablations. Below, we address these point-wise with extensive experiments, collected at https://anonymous.4open.science/r/antigen-rebuttal-31EB/, and hope the reviewer considers recommending acceptance.

Not merging adapters

  • While we never merge the ascent adapters, we still use the learned A matrices for high-LR updates. Not merging is principled: our goal in specialization is not to explicitly induce forgetting, only to improve the target task. Our LRA results show that, at best, merging does not affect target performance, and doing so for AntiGen would be an unnecessary expense.

Improvement over our ablations

  • Appendices B and C present our own ablations in full detail, so that readers can see what each component contributes. They do not contradict the claims in the main paper, which already states that both ablations perform nearly as well as AntiGen (Sections 4.2 and 5.2) before describing what differs (Figure 4). Concretely, our ablations establish the following.
    • The low-rank intervention improves over GRPO at every saturated specialist, and AntiGen does so most reliably (Sections R1–R4).
    • With frozen ascent-derived addresses, AntiGen trains 41% fewer parameters than high-LR LoRA (67.2M against 113.6M in the Countdown setting) while reaching a similar or slightly better accuracy.
    • Consistency. Against frozen random addresses, AntiGen ends higher on Countdown (57.2 ± 2.9 against 53.4 ± 3.9, five seeds) and is the only variant whose end-of-training gain over continued GRPO is significant (+5.9, p = 0.023; random addresses +2.1, n.s.; Section R4).
    • Focal forgetting. Donor log-probability loss is 0.23 nats with the ascent-derived addresses against 0.09 with random ones, and with random addresses donors lose less than unrelated tasks (Section R7). Our revision will state this more clearly in the main text.
  • Regrettably, our appendix contained two errors that the reviewer correctly highlighted: the Appendix C.3 sentence on math tasks was inaccurate (as Fig. 24 shows, math ratios stay within 6% of the specialist's while donor ratios fall by 8–27%), and Fig. 6's donor group omitted one of the five OpenBookQA donors; both are corrected in the revision.

Novelty and related work

  • We thank the reviewer for highlighting these references, and will incorporate them in the final version. We respectfully disagree that they limit our contribution. The objection addresses only the third of the paper's three stated contributions, and none of the cited works studies RL specialists, saturation, or forgetting.
  • Our work contributes (1) the observation, across 16 released specialists in eight domains (Table 1), that specialization leaves non-target performance largely intact, most completely under RL; (2) a characterization of that non-forgetting in a saturated specialist and a study of focal forgetting by low-rank ascent (Table 2); and (3) a new take on specialization that follows from (1) and (2): focusing high-plasticity updates on capacity that is tied to unrelated tasks. Further, from the saturated Countdown specialist, Gate & Up-only tuning at the specialist's learning rate is indistinguishable from continued GRPO, as we show in Section R6.

Q1: Specializing beyond saturation

  • As we show in Sections R1–R3, AntiGen (and our own ablations) continue to improve on GRPO and the RL plasticity baselines even from saturated specialists.
  • On Countdown (Section R1), AntiGen resumes improvement from a 3000-step specialist at which continued GRPO has nearly stopped, and ends well above continued GRPO and random addresses. On DeepScaleR (Section R2), all three variants improve on continued GRPO over three seeds at 0.8B, with AntiGen the most consistent across seeds, and AntiGen improves on every seed at 2B. On OpenBookQA (Section R3), all three variants improve on continued GRPO over three seeds, and only AntiGen's gain on the test split is significant (p = 0.045).

Q2: AI2-RC in the OpenBookQA donor pool

  • The reviewer correctly identifies that AI2-RC is present in the donor pool for OpenBookQA, and may thus cause leakage. We thank the reviewer for this astute observation, and present a rectified result in Section R5.
  • We re-ran the experiment with donors chosen by a rule that never evaluates AI2-RC, excluding both AI2-RC views and the multiple-choice twin of OpenBookQA, with all other settings unchanged. Over the last 800 steps, OpenBookQA accuracy is 72.1 with the AI2-RC-free donors, 72.1 with the submitted donors and 69.8 for continued GRPO; AI2-RC is 78.3, 77.4 and 75.5. The AI2-RC gain is therefore not an artifact of donor selection: it is slightly larger without AI2-RC among the donors. This is not surprising, as we expect performance on the donor tasks to worsen.

Thank you for the detailed response and the new experiments. I have read the response and the results in the anonymous repository. Some of my issues are well addressed but my major concern still remains. The response indicates that the high-learning-rate low-rank update drives the gain and no significant accuracy advantage is over plain LoRA or random addresses. Therefore, I keep my original score.

Reviewer_dtK2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper interestingly flips the common discussion of catastrophic forgetting to consider when it may actually be welcome and how we can utilize it. In particular, this paper studies why small, RL- or SFT-specialized LLMs retain substantial non-target capabilities and and how we can use said targeted forgetting to reclaim the dead weight. First, they find that specialists barely forget anything, that is every specialist still performs well on things it will never be used for. Directly forcing forgetting with Low-Rank Ascent (gradient ascent adapters on donor tasks) removes donor performance but doesn’t reliably help the target. Building on this, they introduce ANTIGEN, which instead reuses the ascent-derived subspaces as a frozen low-rank adapter trained at a much higher learning rate during GRPO. On synthetic arithmetic search (Countdown) and two reasoning settings (DeepScaleR math, OpenBookQA), ANTIGEN improves over continued GRPO and several plasticity baselines.

Critical strengths
  • It is a unique idea. ANTIGEN is an interesting and relatively novel take on capacity reallocation. I don't remember seeing something like it before. Also, while not the main contribution, the work does challenge a common assumption that specialization inevitably hampers non-target performance. This will prompt fun discussion.
  • The narrative is well motivated and explained clearly throughout. The logical story seems sounds.
Critical weaknesses
  • There are very modest margains of improvement. Random A or high-LR LoRA get really close to ANTIGEN’s peak gains in some settings. It makes me wonder if (what seems to be) an incremental benefit is worth it given increased compute costs. The significance of the contribution is weak to me.
  • How strongly can we say "beyond saturation" is really supported. Only Countdown experiments actually test it. And another notable thing about this to me is the countdown control still improving at about +0.7 points per 100 steps. The evidence right now can only really say that ANTIGEN can speed up or maybe extend ongoing specialization on one task, not that it unlocks gains in a model that has stopped improving (which would be the most significant contribution).
Questions for authors

See weaknesses for the most part, but if I had to make a targeted question:

  • How do you explain the small gap between ANTIGEN and the random-address and plain-LoRA variants? Is there evidence, beyond peak accuracy and variance, that the ascent-derived subspaces specifically matter?
1 comment
AuthorsOpenReview ↗

Response to Reviewer dtK2

We thank the reviewer for their thoughtful review of our work, and are happy to see their positive assessment of clarity and soundness. We address each of their concerns point-wise below, with new results collected at https://anonymous.4open.science/r/antigen-rebuttal-31EB/, and we hope that they consider recommending acceptance.

Cost vs. benefit of added computation

  • Across every setting that we evaluate, AntiGen's gain over continued GRPO is significant. We achieve improvements of up to 10 points on Countdown, 2–5 points on DeepScaleR, and 1–2 points on OpenBookQA. For specialists that have otherwise stopped improving, these are not incremental margins. Further, the added compute cost is negligible: the ascent runs once, stops at a loss ceiling after 9–19 adapter steps per donor, and takes about two minutes of a nine-hour run (under 0.5%).
  • AntiGen is also cheaper to train than high-LR LoRA, which performs comparably. For example, in the Countdown setting, AntiGen with the frozen A matrix updates 67.2M parameters (8.9% of the model), whereas plain LoRA at the same rank updates 113.6M (15.1%). Thus, AntiGen is cheaper than this ablation but reliably achieves a similar or slightly better accuracy.
  • Further, the added two-minute cost is a small price to pay relative to random addresses, as AntiGen improves on GRPO more reliably and leads to more focal forgetting than this ablation.

Specializing beyond saturation

  • The reviewer raises important questions on the claim of "specializing beyond saturation". We conduct extensive experiments with long GRPO runs on Countdown, DeepScaleR, and OpenBookQA, and show in Sections R1–R3 that AntiGen does indeed improve on GRPO and other baselines from saturated specialists.
  • Countdown (Section R1). We continue training Qwen3.5-0.8B on Countdown to 3000 steps, after which continued GRPO improves by only about 0.1 points per 100 steps (against 0.7 after the 1600-step specialist of the submission). Even from this specialist, AntiGen resumes improvement within a few hundred steps and ends well above continued GRPO and random addresses, and above high-LR LoRA on average.
  • DeepScaleR (Section R2). We train Qwen3.5-0.8B and 2B on DeepScaleR for 2000 steps, after which continued GRPO improves by only 0.1–0.2 points per 100 steps. From these specialists, AntiGen improves on continued GRPO at both model sizes and for every seed, with the largest gain at 2B, and it remains above the RL plasticity baselines.
  • OpenBookQA (Section R3). We train for 3000 steps, after which continued GRPO is nearly flat (0.08 points per 100 steps). Over three seeds, AntiGen improves significantly on continued GRPO on both the test and validation splits within 1000 steps (p = 0.045 and 0.007); in a longer run to 5000 steps, it keeps improving while continued GRPO and two of the three RL plasticity baselines remain nearly flat.

The benefit of ascent-derived subspaces

  • AntiGen's method of deriving the A matrix by gradient ascent has the following benefits over LoRA and random addresses.
  • Focal forgetting. Averaged over seeds, AntiGen's log-probability loss on the donor tasks is 0.23 nats, 2.5–3.5 times that of the random-address and LoRA containers (0.09 and 0.07), at equal target accuracy. With random or trainable addresses the donors lose less than unrelated probe tasks (donor / probe ratio 0.63 and 0.53); only the ascent-derived addresses direct forgetting toward the donors (Section R7).
  • Parameter efficiency. Frozen ascent-derived addresses reach LoRA's accuracy with 41% fewer trainable parameters. We do not claim that AntiGen's accuracy is significantly above that of the two ablations: at five seeds, those direct tests are not significant (p = 0.62 and 0.13). A readout analysis we add explains why the gap is small: the container's learned update is 78–86% position-invariant, a near-constant write into the residual stream.
  • Reliability at the end of training. AntiGen has the highest final accuracy (57.2 ± 2.9, against 53.4 ± 3.9 for random addresses) and the smallest seed spread, and it is the only variant whose end-of-training gain over continued GRPO is significant (+5.9, p = 0.023; random addresses: +2.1, n.s.; Section R4). From the saturated Countdown specialist, random addresses end close to continued GRPO, while AntiGen ends several points higher (Section R1).

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to reviewer comments

We thank all reviewers for their careful reading and constructive feedback. Our response adds the experiments the reviewers requested, collected at https://anonymous.4open.science/r/antigen-rebuttal-31EB/, and we believe they strengthen the central claims of the paper.

  • Specialization beyond saturation. From specialists whose continued GRPO training has nearly stopped improving, AntiGen resumes improvement on Countdown, on DeepScaleR at two model sizes, and on OpenBookQA, with at least three seeds of AntiGen in each setting, and it is at or above every RL plasticity baseline we evaluate (Sections R1–R3).
  • Reporting. We now show full learning curves and report the mean of the last five checkpoints and the final checkpoint of every method. With five seeds on Countdown, AntiGen's gain over continued GRPO holds under all three evaluations: it is significant for the last five checkpoints and the final checkpoint (p = 0.010 and 0.023), and AntiGen's checkpoint selected on a held-out validation split is, averaged over seeds, above the best checkpoint of every continued-GRPO run (Section R4).
  • What the ascent-derived addresses contribute. Our ablations show that the high-plasticity low-rank update drives the gain over GRPO, while the ascent-derived addresses determine what is forgotten. AntiGen directs forgetting toward the donor tasks (Section R7), trains 41% fewer parameters than LoRA in the same container, and improves on GRPO more reliably than random addresses. We do not claim a significant accuracy advantage over these ablations, and the revision states this explicitly.
  • Donor selection. With donors chosen without reference to AI2-RC, the OpenBookQA and AI2-RC gains are unchanged (Section R5).
  • Related work. We will cite and discuss LoRA+, ReLoRA, asymmetric LoRA and selective-module tuning, relate our ablations to them, and include the Gate & Up-only recipe of Zhu et al. as a baseline, which does not improve on continued GRPO at a saturated specialist (Section R6).

Dear Reviewers,

The authors have responded and haven't received any feedback. Please, read the responses, say whether your concerns were addressed and when appropriate raise your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.