Pith. sign in

REVIEW 4 major objections 4 minor 25 references

Control LLM: Controlled Evolution for Intelligence Retention in LLM

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Control LLM proposes that inserting trainable transformer blocks beside frozen ones and aligning their hidden states with interpolation and a divergence loss lets an LLM learn new skills while retaining old ones.

desk verdict Useful method for mitigating catastrophic forgetting, but the SOTA claim rests on comparisons against models trained from different bases; the controlled same-base evidence is solid. read the letter →

arxiv 2501.10979 v2 pith:JGCKAJED submitted 2025-01-19 cs.LG

Haichao Wei , Yunxiang Ren , Zhoutong Fu , Aman Lunia , Yi-Lin Chen , Alice Leung , Ya Xu This is my paper · ORCID
classification cs.LG
keywords catastrophicforgettingcontinuouspre-trainingsupervisedfine-tuningmodelexpansionhidden-statealignmentinterpolationdivergencelosslargelanguagemodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Control LLM is a training architecture for updating a large language model on new data without catastrophic forgetting. Instead of unfreezing the whole model, it inserts trainable copies of selected transformer blocks beside frozen originals, blends the two branches' hidden states with an interpolation weight, and adds a divergence loss that keeps the trainable branch's representations close to the frozen branch's. The paper reports that this 'learn more, forget less' recipe works for continuous pre-training and supervised fine-tuning alike: on Llama-3.1-8B it lifts math, coding, and Chinese benchmarks while holding MMLU within 4.3 points of the base model, where full fine-tuning loses more than 35 points. If the recipe holds, it offers a cheaper path to specializing models without retraining from scratch.

What carries the argument

The central object is a dual-branch transformer block: for each selected layer, a frozen copy of the original block runs in parallel with a trainable copy, and their hidden states are fused by an interpolation function (linear, dynamic, input-conditioned, progressive, or hard MoE gating). A divergence loss, averaged over layers and weighted per token, penalizes the distance between the two branches' hidden states. This combination is what the paper claims carries the anti-forgetting effect: the interpolation keeps the trainable branch plugged into the frozen model's information flow, while the divergence loss prevents the expanded branch from drifting into representations that erase old skills.

What would settle it

Sweep the divergence weight lambda over a wide range (from zero to very large) on a fixed base model and new-task dataset, and record new-task accuracy alongside MMLU. If no lambda value yields both near-full-fine-tuning new-task accuracy and MMLU within 5 points of the base model, then the claimed 'learn more, forget less' balance does not hold as stated. A simpler check: train on a synthetic task deliberately far from the base distribution, where hidden-state alignment would force the trainable branch to stay close to representations that cannot express the task; if the method fails there, its applicability is limited to tasks near the original distribution.

Watch

Extended reading notes

Core claim

The paper's central claim is that catastrophic forgetting in LLM updates is driven by drift in hidden-state representations, and that drift can be controlled by running every selected layer as two parallel branches: a frozen pre-trained block that preserves old knowledge and a trainable expanded block that absorbs new data, with their outputs combined by interpolation and a per-layer divergence loss that penalizes separation. The authors argue this alignment is what allows the model to 'refine rather than overwrite' prior knowledge. They support the claim with experiments on math (OpenMath2), coding (OpenCoder), and multilingual (Llama-3-SynE) updates, showing that Control LLM matches or exceeds full-parameter tuning on new-task benchmarks while keeping MMLU, ARC, and GPQA near base-model levels, and that the best configuration is Concat expansion with Lerp interpolation and MSE divergence loss.

Load-bearing premise

The load-bearing assumption is that penalizing the trainable branch's hidden states to stay close to the frozen branch's states will not prevent the trainable branch from learning the new task; the paper provides no formal condition for this trade-off, only empirical curves, and the value of the penalty weight lambda is never reported.

Editorial extensions

If this is right

  • Specializing an LLM for math, code, or a new language no longer requires retraining the whole model or replaying the original training data.
  • The approach works in both continuous pre-training and supervised fine-tuning, so a deployed model can be refreshed on new data without a full retraining cycle.
  • Original capabilities degrade by less than 4.3% on MMLU while new-task gains reach +14.4% on Math-Hard and +10% on MBPP-PLUS, a retention profile that standard fine-tuning does not approach.
  • Concat expansion with Lerp interpolation and MSE divergence loss is the recommended default, with the ablation showing that alignment and interpolation together, not either alone, produce the effect.
  • Because only the expanded blocks and interpolators are trained, the method uses substantially less compute and data than full retraining while matching open-source models tuned from the same base.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: A testable consequence follows from the paper's own Lerp8MSE0* ablation: training with alpha=0 and inferring at alpha=0.5 preserves old skills while learning new ones, suggesting the interpolation weight could be tuned at inference time as a retention-versus-adaptation dial without retraining.
  • Editorial extension: If hidden-state alignment is the active mechanism, the same frozen-plus-trainable parallel block design could transfer to vision-language or multimodal continual learning, where representation drift is also implicated in forgetting.
  • Editorial extension: The paper leaves the divergence weight lambda unreported; a natural follow-up is to map the lambda-versus-new-task-accuracy curve to see whether there is a principled operating point, or whether the method's success depends on a narrow range.
  • Editorial extension: The reported production deployment suggests the added parameters from the parallel blocks are acceptable in low-latency serving settings, which is a consequence the paper states but does not quantify.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes Control LLM, a model-upcycling method that adds a trainable expanded transformer block in parallel with each frozen pre-trained block (one every N-1 layers) and fuses their hidden states through interpolation (Lerp, Dlerp, DlerpIn, Plerp, MoE) with an optional divergence loss. The aim is to mitigate catastrophic forgetting during continuous pre-training (CPT) and continuous supervised fine-tuning (CSFT) while still learning new skills. The authors evaluate on math (OpenMath2), coding (OpenCoder), and Chinese multilingual (Llama3-SynE) tasks using Llama-3.1-8B/Instruct, reporting improved new-task performance with much smaller degradation on original capabilities than full-parameter tuning. The paper also includes step-level training curves, an ablation study of interpolation mechanisms, divergence losses, and expansion strategies, and it releases training/evaluation code and models.

Significance. If the central claim is valid, Control LLM would be a practical alternative to full fine-tuning for continual post-training, with useful properties for production settings: retention of broad capabilities, reduced data and compute, and no reliance on replay of original data. The paper has several concrete strengths: it reports training curves at 1k-step granularity, ablates interpolation strategies and divergence losses, compares against full-parameter and partial-parameter tuning, and releases code and models. These assets make the empirical findings checkable. However, the headline claim of state-of-the-art performance among open-source models tuned from the same base model is not supported by the current comparison tables, because key external baselines start from different base checkpoints. The internal same-base comparisons (Full Param Tune, Partial Param Tune, Stack Expansion) are informative and do support the forgetting-mitigation benefit, but the broader SOTA claim needs either new controlled experiments or a substantially narrowed statement.

major comments (4)
  1. [Abstract; Section 4.1.2; Tables 1-3] The claim that Control LLM 'surpasses existing methods and achieves SOTA among open-source models tuned from the same base model' is not established by the reported comparisons. In Table 1, OpenMath2-Llama3.1 is fine-tuned from the non-instruct Llama-3.1-8B base, whereas Control LLM starts from Llama-3.1-8B-Instruct (Section 4.2.1). In Table 2, OpenCoder-8B-Instruct is built on the separately pre-trained OpenCoder base. In Table 3, Llama-3-SynE is derived from Llama-3, not Llama-3.1-8B. Starting from an instruction-tuned model can substantially raise original-capability scores (MMLU, ARC, GPQA) and may also affect new-task performance. The paper needs either same-base external baselines or a revision of the SOTA claim to what the data actually show: superiority over the internally implemented full/partial/stack baselines on the same starting checkpoint.
  2. [Section 3.4, item 5; Tables 1-3] The checkpoint-selection protocol is a source of comparison bias. The text says 'Pick best-performing checkpoint on new tasks' after training 2-6 epochs, while OpenMath2 is reported as a single released checkpoint and the paper explicitly notes that OpenMath2 uses checkpoint averaging. Selecting the best new-task checkpoint for Control LLM, but not applying the same selection to external baselines, can inflate the reported margin. The authors should either apply the same best-checkpoint selection to every comparison model or report results at a fixed training budget with a clear protocol.
  3. [Section 3.4, Eq. (1); Section 4.2.3] The divergence-loss weight lambda in L = L_task + lambda * L_divergence is never reported, despite being a central hyperparameter of the method. The ablation names such as Lerp8MSE0 suggest lambda=0 in one configuration, but no lambda value is given for the main results. Since the authors release code, this is partially mitigated, but the paper should state the lambda values used for each experiment group, together with the divergence metric (MSE vs cosine) for each reported row.
  4. [Section 3.1; Appendix B] The paper's explanation that hidden-state alignment is the cause of catastrophic-forgetting mitigation is presented as a finding, but the evidence is correlational: probing analogous sentences and comparing nearest-neighbor distances shows that well-performing configurations have better-aligned hidden states. The training objective explicitly enforces such alignment, so the observed correlation is expected by construction. This does not invalidate the method, but the causal language in the contributions ('Highlight the role of hidden-state alignment in retaining prior knowledge') should be softened to a hypothesis, or supported by an intervention study that varies alignment independently of the training objective.
minor comments (4)
  1. [Section 4.1.2] The sentence 'We also compare to open-source models trained on similar datasets and same base model' is inaccurate for the three listed models, as none of them is trained from the same base checkpoint as Control LLM. This should be corrected to avoid confusing readers.
  2. [Tables 1-3] The table captions state 'Avg.' and the text says results are 'size-weighted averages within each task group, along with the overall mean of these group averages,' but the reader cannot verify which columns are averaged and how the size weights are computed. Please include the exact averaging formula in a footnote or appendix.
  3. [Section 4.2.3] The ablation table uses compact names such as Lerp8MSE0, Lerp8MSE0*, Lerp8MSE0*M, and Lerp8MSEα without a key. A one-sentence explanation of each suffix (e.g., '0' means alpha=0 during training, '*' means inference at alpha=0.5, 'M' means merged weights) would greatly improve readability.
  4. [Throughout] There are several typos and formatting issues: 'harvness' for 'harness', 'Con-trol LLM' line breaks, and inconsistent use of 'full-parameter' vs 'full parameter'. These should be cleaned up before publication.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical evaluations against external benchmarks, and no fitted parameter or cited theorem is renamed as a prediction.

full rationale

The paper is an empirical method paper: Control LLM combines frozen pretrained blocks, trainable expanded blocks, interpolation, and a divergence regularizer, then measures held-out benchmark scores (Math-Hard, MBPP-Plus, C-Eval, MMLU, etc.). The training objective L = L_task + lambda L_divergence directly enforces hidden-state closeness, so one might say 'alignment preserves original capabilities' is partly by construction; however, the paper does not present this as a derived prediction, and the decisive outcomes (new-task learning, retention trade-offs) are external measurements that could have failed. Ablations exhibit genuine trade-offs (e.g., Dlerp without divergence loss retains without explicit MSE; MoE gating underperforms; Lerp without divergence loss forgets more), so the results are not forced by the loss definition alone. The claimed SOTA among same-base models is contestable because Tables 1-3 compare models initialized from different checkpoints, but that is an experimental-validity issue, not circularity. No load-bearing self-citations or imported uniqueness theorems appear; prior work (LLaMA Pro, LoRA, EWC) is used for baselines and context, not to justify the central claim. Therefore no circular step meets the evidentiary bar.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical entities or unobserved quantities. The 'expanded block' and 'interpolator' are architectural components whose existence is directly instantiated in code and checkpoints, not speculative postulates.

free parameters (5)
  • Interpolation weight alpha (Lerp) = 0.5 (frozen in final runs)
    Chosen by hand and fixed to 0.5 after ablations (Section 4.2.3); central to the blending of frozen and expanded hidden states.
  • Number of expanded layers N = 8 (production), 16 (open-source)
    Selected via ablation (Table 5); affects capacity and compute.
  • Divergence loss weight lambda = Not reported in text
    The training loss includes lambda * L_divergence (Section 3.4) but the value is omitted, making exact reproduction difficult.
  • Divergence metric = MSE (default)
    Chosen empirically; cosine also works but degrades with prolonged training (Section 4.2.3).
  • Dlerp gating parameters (W, b) = learned during training
    Used in dynamic interpolation when Dlerp is selected; not used in final Concat-Lerp default.
assumptions (3)
  • domain assumption Hidden-state alignment, measured by cosine/Euclidean distance between expanded and frozen blocks, is a valid proxy for knowledge retention.
    Used throughout Section 3.1 and Figure 3 to motivate the method; no formal proof is given that alignment causes retention.
  • domain assumption The divergence loss can be added to the task loss without suppressing new-task learning; the balance is controlled by lambda.
    Training objective in Section 3.4; paper provides no theory for the trade-off.
  • domain assumption Fine-tuning on task-specific data without replay is a valid test of catastrophic forgetting.
    Section 4.1.1 explicitly omits English data in CPT to test CF without replay.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Control LLM: Controlled Evolution for Intelligence Retention in LLM." pith.science (2026). https://pith.science/paper/JGCKAJED

@misc{pith2026250110979,
  author       = {Pith},
  title        = {Pith review of: Control LLM: Controlled Evolution for Intelligence Retention in LLM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JGCKAJED}},
  note         = {Machine review of arXiv:2501.10979}
}
abstract

Large Language Models (LLMs) demand significant computational resources, making it essential to enhance their capabilities without retraining from scratch. A key challenge in this domain is \textit{catastrophic forgetting} (CF), which hampers performance during Continuous Pre-training (CPT) and Continuous Supervised Fine-Tuning (CSFT). We propose \textbf{Control LLM}, a novel approach that leverages parallel pre-trained and expanded transformer blocks, aligning their hidden-states through interpolation strategies This method effectively preserves performance on existing tasks while seamlessly integrating new knowledge. Extensive experiments demonstrate the effectiveness of Control LLM in both CPT and CSFT. On Llama3.1-8B-Instruct, it achieves significant improvements in mathematical reasoning ($+14.4\%$ on Math-Hard) and coding performance ($+10\%$ on MBPP-PLUS). On Llama3.1-8B, it enhances multilingual capabilities ($+10.6\%$ on C-Eval, $+6.8\%$ on CMMLU, and $+30.2\%$ on CMMLU-0shot-CoT). It surpasses existing methods and achieves SOTA among open-source models tuned from the same base model, using substantially less data and compute. Crucially, these gains are realized while preserving strong original capabilities, with minimal degradation ($<4.3\% \text{on MMLU}$) compared to $>35\%$ in open-source Math and Coding models. This approach has been successfully deployed in LinkedIn's GenAI-powered job seeker and Ads unit products. To support further research, we release the training and evaluation code (https://github.com/linkedin/ControlLLM) along with models trained on public datasets (https://huggingface.co/ControlLLM) to the community.

Figures

Figures reproduced from arXiv: 2501.10979 by the authors.

Figure 1
Figure 1. Comparison: Ours vs SOTA Llama-tuned models. FLOPs (Brown et al., 2020; Dubey et al., 2024). These computational demands make full retraining impractical for many. While pre-trained LLMs provide broad utility, they often lack domain-specific skills. Improving these models via continuous pre-training (CPT) offers a practical alter￾native to full retraining. Post-training methods—such as supervised fine-tuning (SFT), … view at source ↗
Figure 2
Figure 2. [Result] Comparison of CF - our method vs others on open-source datasets: (left) OpenMath, (right) OpenCoder [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. [Why] Hidden State Alignment Comparison: Best Alignment(Control LLM) vs Worst Alignment(Full-Parameter Tuning) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: [How] Control LLM Architecture. (a) Expanded blocks added every N − 1 layers connect to frozen blocks via interpolators. (b) Interpolators align hidden-states to produce final representations. (c) Different interpolation strategies are explored. Dynamic Linear Interpol…
Figure 5
Figure 5. Figure 5: illustrates three expansion strategies for each (N − 1)th layer: • Concat: Adds a parallel “side-car” for dual-branch blend￾ing, initialized as a copy of pretrained transformer block. • Stack: Stacks copied new layers (adapted from Wu et al. (2024)) above the original,…
Figure 6
Figure 6. Figure 6: [OpenMath + Llama-3.1-8B-Instruct] Benchmark comparison of training methods. (a) Full Parameter Tuning. (b) Partial Parameter Tuning: Freeze all except 1 of every 4 transformer layers (8 total). (c) Stack Expansion: Add 8 transformer layers (1 per 4) while freezing ori…
Figure 7
Figure 7. Figure 7: [OpenCoder SFT Phase2 + Llama-3.1-8B-Instruct] Benchmark comparison of training methods. (a) Full Parameter Tuning. (b) Partial Parameter Tuning: Freeze all except 1 of every 4 transformer layers (8 total). (c) Stack Expansion: Add 8 transformer layers (1 per 4) while …
Figure 8
Figure 8. Figure 8: [OpenMath2 + Llama-3.1-8B-Instruct] Comparison of various alignment strategies and their impact on model performance. (a) Pre-trained Model: Hidden states from [Pretrained] and [Expanded] blocks are identical for all layers. 1) Semantic Stability: Analogous sentences e…
Figure 9
Figure 9. Figure 9: [OpenMath2 + Llama-3.1-8B-Instruct] Comparison of benchmark results of different ablation study settings every 1K steps during training. (a) Lerp Interporation Strategy with MSE loss. (b) Lerp Interporation Strategy without Divergence Loss. (c) Dlerp Interporation Stra…
Figure 10
Figure 10. Figure 10: [OpenCoder + Llama-3.1-8B-Instruct] Comparison of benchmark results of different ablation study settings during training. (a) Lerp Interporation Strategy with MSE Loss. (b) Lerp Interporation Strategy with MSE Loss. Trained 3X longer than (a) (c) Hybrid Expansion Stra…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages

  1. [1]

    Semantic Stability: Analogous sentences exhibit high similarity in both cosine and Euclidean distances

    work page

  2. [2]

    The model retains original capabilities with no CF

    Layer-wise Clustering: Hidden states form distinct clusters for each layer. The model retains original capabilities with no CF. • (b) Control LLM (Lerp + MSE Divergence Loss): Both Semantic Stability and Layer-wise Clustering are preserved. Distances between [Pretrained] and [Expanded] are minimal, yielding strong CF mitigation. MMLU drops slightly from 0...

    work page

  3. [3]

    https://huggingface.co/datasets/lighteval/MATH

    work page

  4. [4]

    https://huggingface.co/datasets/lighteval/MATH-Hard

    work page

  5. [5]

    https://huggingface.co/datasets/openai/gsm8k

    work page

  6. [6]

    https://huggingface.co/datasets/google-research-datasets/mbpp/viewer/sanitized

    work page

  7. [7]

    https://huggingface.co/datasets/evalplus/mbppplus

    work page

  8. [8]

    https://huggingface.co/datasets/openai/openai_humaneval

    work page

Show all 25 references
  1. [9]

    https://huggingface.co/datasets/evalplus/humanevalplus

  2. [10]

    https://huggingface.co/datasets/ceval/ceval-exam

  3. [11]

    https://huggingface.co/datasets/haonan-li/cmmlu

  4. [12]

    https://huggingface.co/datasets/meta-llama/Llama-3.1-8B-Instruct-evals/viewer/Llama-3.1-8B-Instruct-evals__arc_challenge__details

  5. [13]

    https://huggingface.co/datasets/meta-llama/Llama-3.1-8B-Instruct-evals/viewer/Llama-3.1-8B-Instruct-evals__gpqa__details

  6. [14]

    https://huggingface.co/datasets/meta-llama/Llama-3.1-8B-Instruct-evals/viewer/Llama-3.1-8B-Instruct-evals__mmlu__0_shot__cot__details

  7. [15]

    https://huggingface.co/datasets/meta-llama/Llama-3.1-8B-Instruct-evals/viewer/Llama-3.1-8B-Instruct-evals__mmlu_pro__details

  8. [16]

    https://huggingface.co/datasets/meta-llama/Llama-3.1-8B-evals/viewer/Llama-3.1-8B-evals__bbh__details

  9. [17]

    https://huggingface.co/datasets/meta-llama/Llama-3.1-8B-Instruct-evals/viewer/Llama-3.1-8B-Instruct-evals__math__details

  10. [18]

    https://huggingface.co/datasets/meta-llama/Llama-3.1-8B-Instruct-evals/viewer/Llama-3.1-8B-Instruct-evals__math_hard__details

  11. [19]

    https://huggingface.co/datasets/meta-llama/Llama-3.1-8B-Instruct-evals/viewer/Llama-3.1-8B-Instruct-evals__gsm8k__details

  12. [20]

    https://huggingface.co/datasets/meta-llama/Llama-3.1-8B-Instruct-evals/viewer/Llama-3.1-8B-Instruct-evals__mbpp__details

  13. [21]

    https://huggingface.co/datasets/meta-llama/Llama-3.1-8B-Instruct-evals/viewer/Llama-3.1-8B-Instruct-evals__mbpp_plus__details

  14. [22]

    https://huggingface.co/datasets/meta-llama/Llama-3.1-8B-Instruct-evals/viewer/Llama-3.1-8B-Instruct-evals__human_eval__details

  15. [23]

    https://huggingface.co/datasets/meta-llama/Llama-3.1-8B-Instruct-evals/viewer/Llama-3.1-8B-Instruct-evals__human_eval_plus__details

  16. [24]

    https://github.com/EleutherAI/lm-evaluation-harness/blob/main/lm_eval/tasks/ceval/_default_ceval_yaml

  17. [25]

    Evaluation datasets, abbreviations, data source, and prompt templates

    https://github.com/EleutherAI/lm-evaluation-harness/blob/main/lm_eval/tasks/cmmlu/_default_template_yaml 15 Control LLM: Controlled Evolution for Intelligence Retention in LLM Table 7. Evaluation datasets, abbreviations, data source, and prompt templates. ∗ - input final promp...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.