acceptodds
Under review as a conference paper at ICLR 2027

SIFT: Boundary-Free Online Adaptation via Semantic Isolation and Functional Testing

Abstract

Boundary-free online continual learning aims to adapt large language models to a stream of evolving data without access to task identities or boundaries. Without this supervision, the learner must decide from the stream alone when incoming data can safely reuse existing adaptation capacity and when it requires separate capacity. Sharing one set of adaptation parameters across the stream causes interference and forgetting, whereas allocating separate parameters typically relies on externally provided task boundaries and can increase parameter counts unnecessarily. We present SIFT (Semantic Isolation and Functional Testing), a framework for boundary-free online adaptation that keeps the language model backbone frozen and organizes the stream into mature semantic modules and provisional modules. Because nearby regions can still require conflicting updates, semantic proximity alone cannot certify that adaptation capacity can be shared, so SIFT defers the decision until functional evidence is available. It routes examples to mature modules by cosine distance, isolates uncertain ones into provisional modules that inherit a parent prompt but adapt independently, and consolidates a provisional module only when a radius-based compatibility check and a sentinel-based functional validation agree. We conduct extensive experiments using four instruction-tuned backbones from the Llama and Qwen families, spanning 0.6B to 7B parameters. SIFT achieves up to 21.96% relative improvement in final accuracy over the strongest baseline while substantially reducing forgetting.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
continual learning, parameter-efficient adaptation, catastrophic forgetting mitigation

Reviews

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

This paper proposes SIFT, a framework for boundary-free online continual learning in large language models. The backbone remains frozen, while a collection of soft-prompt modules is organized into mature and provisional modules. Incoming examples are routed according to source-representation similarity. When an example cannot be safely assigned to a mature module, a provisional branch inherits its parent prompt and adapts independently. Once the branch has accumulated sufficient evidence, SIFT combines a radius-based semantic compatibility check with a functional validation step on a retained sentinel set to decide whether the branch should be merged or maintained separately. Experiments on six NLP datasets and four Llama/Qwen backbones show improvements over several continual-learning baselines, particularly in final accuracy and backward transfer.

Critical strengths
  1. The paper addresses a relevant problem in online continual learning: semantic similarity alone does not guarantee that two data regions can safely share adaptation parameters. The distinction between semantic compatibility and functional compatibility is meaningful, and the provisional-module mechanism provides a clear way to isolate potentially interfering updates before deciding whether to merge them.

  2. The proposed framework is evaluated on heterogeneous NLP tasks and multiple instruction-tuned backbones. The ablation study supports the contribution of provisional modules, parent-prompt initialization, and functional validation. The comparison with shared-prompt and shared-adapter baselines also illustrates the potential value of isolating updates rather than continually overwriting a single adaptation state.

Critical weaknesses
  1. The robustness of the main empirical conclusions is not sufficiently established. The main results are reported as point estimates without clearly presenting uncertainty across random seeds or alternative stream orders. Moreover, the gains vary considerably across backbones: the improvement over F2OCL is substantial on Llama-3.2-1B-Instruct but only 0.15 percentage points in final accuracy on Llama-3.2-3B-Instruct. Since the benchmark consists of a single-pass stream of six heterogeneous datasets, it is important to establish whether the reported gains persist under different stream orders and random initializations.

  2. The resource budget and the claim of operating without replay need clearer qualification. The implementation reserves up to 384 mature modules and 128 provisional modules, each with a 100-token soft prompt, and retains a global budget of 384 sentinel examples for functional validation. Although the sentinels are not used for gradient updates, the method still stores historical examples and incurs additional validation and module-management costs. The paper does not sufficiently establish how these costs compare with the memory, trainable-parameter, and inference budgets of the baselines. Parameter- or memory-matched comparisons, together with latency and storage measurements, would help determine whether the performance gains justify the additional capacity.

Questions for authors
  1. Are the main results stable across multiple random seeds and alternative dataset orders? In particular, could the authors report uncertainty estimates for the final-accuracy differences, given the small improvement on Llama-3.2-3B-Instruct?

  2. How do the total trainable parameters, prompt-storage requirements, sentinel memory, and inference overhead compare with those of the baselines? Since SIFT retains historical examples for functional validation, could the authors clarify the precise meaning of the claim that the method operates without replay and provide comparisons under matched resource budgets?

  3. How sensitive are the results to the maximum number of mature modules, the provisional-module capacity, and the sentinel budget? Would a smaller module budget preserve most of the reported performance gains?

1 comment

Dear Reviewer Uznc,

Please carefully review the authors' rebuttal and respond to each of your concerns. If any concerns remain unresolved, clearly specify what is missing and request concrete clarifications or additional evidence.

Bests,

Your AC.

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

This paper studies boundary-free, single-pass, replay-free online continual learning for LLMs. It freezes the backbone and routes examples to soft-prompt modules by nearest-centroid distance: hits update in place, misses spawn an isolated provisional branch. Candidates are later fused or promoted via a semantic overlap check plus a sentinel-based functional check. Across 4 backbones and 6 datasets, SIFT improves final accuracy while reducing forgetting.

Critical strengths
  1. The motivation is grounded in direct empirical evidence rather than assumption. The paper shows that even within dataset updates can hurt prior performance, undermining the common belief that semantic similarity alone predicts safe parameter sharing.

  2. The branch and defer design lets the model warm start new capacity from existing knowledge and postpone the share or isolate decision until real evidence is available, so a wrong routing choice costs an extra branch rather than corrupting what's already learned.

Critical weaknesses
  1. The paper argues similarity cannot decide when to share, but never compares against a learned router or a small replay buffer. Even its own ablation undercuts the claim: with fusion removed entirely, stage accuracy stays at 42.00 vs 42.12, so the checks mainly help final accuracy and forgetting, not the core sharing decision they're meant to justify.

  2. The safety check rests on roughly one example per module (384 examples for up to 384 modules), too little to detect damage to old data, and merging overwrites the parent prompt while testing only the parent. Meanwhile mature modules still absorb similar examples with unchecked in-place updates—the very interference the paper's own 35.6% figure exposes

Questions for authors
  1. Removing fusion barely changes stage accuracy, 42.00 vs 42.12 for the full model. Does that suggest the gain comes from keeping separate branches rather than from the checks and fusion you present as the main contribution?
  2. Section 3 reports that 35.6% of updates within the same dataset hurt earlier data. Mature prompts are still updated in place with no check at all. Did you measure this and find it negligible, or is it simply left unaddressed?
  3. The sentinel budget of 384 is shared across as many as 384 modules, so some modules may get only a single example. Can Eq.15 still detect a harmful merge in that case, and how many modules were actually created?
  4. Eq.15 tests the candidate only on the parent's sentinels, while Eq.16 discards the parent prompt and keeps the candidate. Why not test the candidate's own region, or keep both prompts?
  5. GRID scores below vanilla prompt tuning on two backbones and L2P trails by about 18 points. Can you give the prompt length, pool size and learning rate used for these baselines, and confirm they follow each method's original recommended settings?
2 comments

Dear Reviewer keFB,

Please carefully review the authors' rebuttal and respond to each of your concerns. If any concerns remain unresolved, clearly specify what is missing and request concrete clarifications or additional evidence.

Bests,

Your AC.

Thank you for the detailed responses and additional experiments. The clarifications have addressed some of my concerns. However, I still have the following questions:

  1. Q2 : The new experiment demonstrates that direct mature updates improve current-task performance, but does not evaluate their impact on previously learned tasks. Since the paper highlights within-task interference as a key motivation, I remain concerned that such interference is not explicitly addressed during mature module updates.

  2. Q5: Thank you for clarifying the baseline configurations. However, SIFT uses 39 mature modules at the final stage, whereas L2P and GRID are limited to a pool size of 7. Could the authors discuss whether this capacity difference affects the fairness of the comparison, or provide results under comparable parameter budgets?

  3. Implicit task structure in the training stream: Although task identities and boundaries are not provided, the training stream appears to process samples from each dataset sequentially, rather than interleaving samples across datasets. If so, this may introduce implicit task structure, as consecutive samples from the same dataset tend to be semantically similar, potentially benefiting SIFT's semantic routing and module management. Could the authors clarify whether the training data are indeed organized in this way, and discuss how SIFT would perform if samples from different tasks were fully interleaved during training? Demonstrating strong performance under such settings would further strengthen the paper's contribution and its claim of boundary-free adaptation.

I appreciate the authors' efforts in addressing these concerns, but I believe these points remain important for evaluating the method's effectiveness and generalizability.

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

This paper proposes SIFT (Semantic Isolation and Functional Testing), a framework for boundary-free online continual learning of large language models without access to task identities or task boundaries. It combines semantic routing, isolated prompt adaptation, and functional validation to balance learning new knowledge and preventing catastrophic forgetting. Experiments across multiple datasets and LLM backbones demonstrate improvements over existing baselines.

Critical strengths

The empirical evaluation is solid, covering multiple LLM backbones and datasets, with comprehensive comparisons against existing baselines and ablation studies demonstrating the effectiveness of the proposed components.

Critical weaknesses

First, the paper argues that semantic similarity alone cannot guarantee safe parameter sharing, yet SIFT directly updates mature modules based solely on semantic similarity when it doesn't trigger the provisional prompt generation. This raises concerns about whether the proposed method fully addresses the interference problem it identifies. Second, MoE-based approaches can also perform automatic expert routing without task identities. Comparisons between the online version of MoE and SIFT would help clarify SIFT's novelty and advantages.

Questions for authors

See weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.