Pith. sign in

REVIEW 3 major objections 6 minor 54 references

GRID: Scaling Task-Agnostic Inference in Continual Prompt Tuning

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read GRID claims forgetting in task-agnostic continual prompt tuning is mostly a decoding-space failure: constraining output to each task's remapped labels nearly halves backward-transfer loss and cuts prompt memory to one third.

desk verdict GRID is a genuinely useful combination of constrained decoding and gradient-guided prompt compression, but its 'task-agnostic' claim is weaker than advertised because the label-identification step is handed the candidate label set at test time. read the letter →

arxiv 2507.14725 v4 pith:7L6WMI2H submitted 2025-07-19 cs.LG cs.AI

Anushka Tiwari , Sayantan Pal , Rohini K. Srihari , Kaiyi Ji This is my paper · ORCID
classification cs.LGcs.AI
keywords continuallearningprompttuningtask-agnosticinferencebackwardtransferconstraineddecodingcompressiongradient-guidedselectionlargelanguagemodels
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

The paper is trying to establish that the severe forgetting seen in prompt-based continual learning under task-agnostic inference is mainly a decoding-space failure rather than a prompt-retention failure. Once task identifiers are unavailable for prompt selection, the shared prompt pool still holds the earlier tasks' knowledge, but the model drifts: it emits labels from the wrong task or hallucinates tokens outside any seen label set. GRID attacks this with output-space-aware decoding, masking generation to the semantically remapped label vocabulary of the task inferred for each input, and with gradient-guided compression that merges redundant prompts so the pool stays bounded. The reported result is that average backward transfer on T5-large improves from about $-0.67$ (ProgPrompt) and $-0.55$ (SHLPT) to $-0.35$ while average accuracy holds at 79.2 percent and prompt memory drops from 600 KB to 200 KB.

What carries the argument

The load-bearing object is the decoding mask, $M_j = \mathbf{1}[v_j \in \mathcal{L}_i]$, a binary mask over the vocabulary that restricts the softmax to tokens in the label set $\mathcal{L}_i$ of the task inferred for input $x$, so generation cannot leave the remapped label space. The label set is produced by a two-stage identification module: rule-based heuristics that match label tokens and input structure to known task templates, and, when those fail, zero-shot task classification with a lightweight generative model that also performs semantic label remapping. Memory control is carried by gradient-guided prompt selection: each prompt $p_j$ in the frozen pool is scored by $g_j = \frac{1}{|D_t|} \sum_{(x,y) \in D_t} \|\nabla_{p_j} L_t\|_2$ on the new task's data, prompts with $g_j > \mu_g + \alpha \sigma_g$ are retained, and the rest are merged into $p_{\mathrm{agg}} = \sum w_j p_j$ with weights $w_j \propto g_j$. In the ablation, the decoding mask is the primary driver of backward-transfer gains, while the gradient scorer mainly delivers scalability.

What would settle it

Measure the label-identification module's accuracy in isolation on the 15-task benchmarks: feed each test input through the rules-plus-zero-shot pipeline and compare the inferred task type and label set to ground truth. If that accuracy is materially below perfect yet BWT stays near $-0.35$ when wrong label sets are deliberately injected, such as a permuted mapping, the proposed mechanism is not the cause of the gains. The cleaner experiment is an ablation that swaps the module for ground-truth task IDs versus a deliberately broken label map: the central claim predicts BWT should collapse in the broken-map condition.

Watch

Extended reading notes

Core claim

GRID's central claim is that task-agnostic forgetting in continual prompt tuning is dominated by uncontrolled output spaces: with no task ID to route prompts, unconstrained decoding over a growing prompt pool produces label drift and hallucinated labels, and this is what drags down backward transfer. The framework establishes that constraining each decoding step to the label set of the task type inferred for that input, after normalizing non-descriptive labels such as $\{0,1\}$ and 'choice1/choice2' into consistent semantics, largely recovers the lost retention: in the ablation, removing constrained decoding while keeping compression drops BWT from $-0.35$ to about $-0.69$ on T5-large, nearly matching the baselines. The companion mechanism, gradient-guided prompt selection, keeps the prompt pool bounded: prompts whose average gradient norm on the new task falls below $\mu_g + \alpha \sigma_g$ are merged into a single gradient-weighted aggregate, cutting memory from 600 KB to 200 KB with minimal BWT cost. The paper further shows the gains hold across encoder-decoder and decoder-only backbones (T5, Qwen, LLaMA) and partly extend to open-ended generation.

Load-bearing premise

The load-bearing premise is that the label-identification module, rule-based heuristics backed by a zero-shot generative model, infers the correct task type and remapped label set for every test input; if it picks the wrong label space, constrained decoding locks the model onto the wrong vocabulary and performance collapses, and the paper reports no accuracy or failure analysis for this step.

Editorial extensions

If this is right

  • Task-agnostic deployment of prompt-tuned models becomes practical: a model can answer mixed inputs from 15 past tasks with no task identifier while holding about 79.2 percent average accuracy on T5-large, with similar levels on T5-3B, Qwen-3-4B, and LLaMA-2-7B.
  • Backward transfer roughly doubles in quality: average BWT improves from $-0.67$ (ProgPrompt) and $-0.55$ (SHLPT) to about $-0.35$ on T5-large, a relative improvement exceeding 50 percent across all seq-to-seq variants.
  • Prompt memory becomes effectively bounded and cheaper at inference: retaining about 5 prompts instead of 15 cuts memory from 600 KB to 200 KB and speeds up inference by 30–50 percent on short inputs while wall-clock training time stays comparable (about 27 vs 24 hours on an A100).
  • The retention benefit is not an artifact of fixed label spaces: on open-ended SuperNI tasks where constrained decoding cannot apply, GRID still improves average accuracy and BWT, indicating that prompt compression alone reduces cross-task interference.
  • Task order matters and interacts with the mechanism: the hard-to-easy order (L5) yields the least negative BWT under GRID, a pattern consistent across model scales.

Reading between the lines

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

  • The ablation isolates constrained decoding as the near-sole driver of the retention gains ($-0.35$ vs $-0.69$ without it), which suggests the headline result is fundamentally about output-space control: other prompt-based or parameter-efficient methods that restrict generation to a per-task normalized label vocabulary could capture similar backward-transfer gains without GRID's specific prompt mac
  • The reported redundancy statistic for low-gradient prompts (average cosine similarity at or above 0.87, Euclidean radius below 0.45) is presented as an observation; read as a hypothesis, it implies gradient-norm scoring could serve as a general pruning criterion for other continual parameter-efficient methods, such as adapter or LoRA stacks, with no retraining.
  • Because the paper does not report the accuracy of its label-identification step, the mechanism's reliability ceiling is unknown; a direct accuracy measurement of task-type and label-set inference on the benchmark tasks would settle how far the constrained-decoding assumption can be pushed before it becomes the bottleneck.
  • On the negative-transfer sequences, forward transfer is mixed, with GRID trailing ProgPrompt on T5-base and T5-large, which suggests the compression step can discard transferable signal when tasks are dissimilar; a natural extension would be a similarity-aware fallback that skips merging when the low-gradient cluster is diverse.
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

3 major / 6 minor

Summary. The paper proposes GRID, a continual prompt tuning framework for task-agnostic inference, where task identifiers are unavailable at test time and the prompt pool has bounded capacity. GRID combines (i) representative sample selection via K-means, (ii) a hierarchical label identification module that normalizes label semantics, and (iii) constrained decoding over the remapped label set, together with (iv) gradient-guided prompt selection and aggregation that merges low-information prompts into a single representation. Experiments on six long-sequence orders, negative-transfer benchmarks, and a new Math-Tool benchmark across T5/Flan-T5, Qwen, and LLaMA backbones report improved backward transfer, competitive or improved accuracy, and reduced prompt memory (about 200 KB vs. 600 KB for ProgPrompt). The ablation in Table 6 shows that constrained decoding is the primary driver of the BWT gains, while gradient-based selection mainly reduces memory.

Significance. If the claims hold, GRID addresses a real and underexplored gap: prompt-based continual learning under task-agnostic inference with bounded memory. The experimental scope is broad (six backbones, multiple task orders, ablations, a new benchmark), and the authors are honest in attributing the BWT improvements primarily to constrained decoding (Section 6.6, Appendix C.8). The paper ships code and includes an ablation of each component, which is commendable. However, the central claim of task-agnostic inference is only partially verified: the main mechanism (constrained decoding) depends on per-instance candidate label sets that are provided in the benchmark protocol, and the label identification module's accuracy is never measured. The absence of variance reporting also limits the strength of the comparisons. The significance is therefore conditional on closing this gap.

major comments (3)
  1. [Section 4.1 (item 2) and Definition 1] Constrained decoding, which Table 6 identifies as the primary driver of BWT gains, requires a candidate label set Yi at test time. In the benchmark protocol, Yi is exactly the label space of the task that generated the input x, so the method is given task-specific output-space metadata that is not excluded by Definition 1. This is a form of task knowledge: knowing the candidate label set constrains the decoding to that task's vocabulary. The paper reports no accuracy, confusion, or failure analysis for the rule-based/Phi-3.5 label identification step, and the ablation removes constrained decoding as a whole rather than isolating the label identification component. The central claim of task-agnostic inference is therefore unverified in the setting where Yi is unknown or noisy. Please either (a) add an evaluation of label identification accuracy (oracle Yi vs. predicted Yi), report error rates of the rule and LLM components, and measure downstream BWT/accuracy under misidentification; or (b) explicitly reframe the contribution as task-agnostic prompt selection given per-instance output-label spaces, which is a weaker but accurate statement.
  2. [Tables 3, 4, 5, 14, 21, 22 and Section 5] The results are reported as averages 'over three runs' but no standard deviations, confidence intervals, or significance tests are provided. Several key comparisons are close, e.g., Table 3 (L6: GRID 75.50 vs. SHLPT 76.20), Table 22 (T5-large NT2: GRID 74.84 vs. ProgPrompt 73.63, and NT1: GRID 72.29 vs. ProgPrompt 73.57), and Table 5 (Qwen accuracy: GRID 53.63 vs. ProgPrompt 54.85). Without variance information, it is impossible to judge whether these differences are meaningful or within run-to-run noise, especially since the headline BWT claims rely on small absolute improvements in some settings. Add standard deviations or confidence intervals for at least the main tables, and state the number of seeds in captions.
  3. [Appendix C.2 (Forgotten Task Count definition)] The formal definition of Forgotten Task Count is flawed: a task Ti is considered forgotten at step t if a(t)_i < τ·min_j a(0)_j, where the threshold is the same for every task (the minimum standalone accuracy across all tasks) rather than the task's own standalone accuracy. Under this definition, a naturally high-accuracy task will be counted as forgotten after a small drop, while a low-accuracy task may never be counted even after catastrophic failure. The FTC numbers reported in Section 6.3 (13.8 vs. 78.7 vs. 64.0) are therefore not a reliable measure of forgetting. Please correct the definition to use a per-task threshold (e.g., τ·a(0)_i) or explain the intended metric and recompute all FTC values.
minor comments (6)
  1. [Table 3] The 'GRID*' entry in Table 3 uses an asterisk that is never explained in the text or footnote.
  2. [Table 6] The footnote-like sentence 'Amazon had the lowest standalone accuracy (∼0.50); setting τ=0.4 yields an absolute threshold of 0.20' appears without context; clarify whether this is a footnote to the FTC definition and why Amazon is singled out.
  3. [Section 4.1] The notation Yi (candidate label set) is introduced in the prose but not defined in the problem setup of Section 3.1; please add a formal definition.
  4. [Section 4.2] The claim that low-gradient prompts are 'highly redundant (average cosine similarity ≥0.87, Euclidean radius R <0.45)' introduces R without definition; specify how R is computed.
  5. [Appendix C.8.4 / Table 27] The GPU time '84:72' in Table 27 appears to be a typo; 72 minutes in the seconds field is inconsistent with the other entries, which are in h:mm format.
  6. [Figure 2] Figure 2 is dense and mixes pipeline stages with notation; consider simplifying the figure and moving some components to an appendix for readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GRID's backward-transfer numbers are measured benchmark outcomes, and the main mechanism (constrained decoding) is ablated rather than derived from its inputs.

full rationale

GRID's central claim, substantially enhanced backward transfer, is supported by measured accuracy and BWT tables (Tables 3, 4, 6, 14, 27) rather than by an equation that constructs the result from fitted parameters. The ablation study (Section 6.6, Tables 6 and 27) identifies constrained decoding as the primary contributor, and removing it returns BWT to approximately ProgPrompt levels; that is a component attribution, not a circular derivation. The label-identification module (Section 4.1, item 2) receives a candidate label set Yi and remaps labels via rules or zero-shot Phi-3.5, and no error rate for this step is reported; this is a load-bearing assumption for the task-agnostic claim, but it is an omitted measurement rather than a circular reduction, because the constrained mask only restricts the vocabulary and does not by itself select the correct label. The prompt-selection threshold alpha (Eq. 2) is tuned on the same benchmarks (Table 13), but the paper reports stability over alpha in [0.6, 0.8], and the sensitivity table is standard hyperparameter analysis, not a fitted parameter renamed as a prediction. The only self-citation (Tiwari and Ji, 2026) appears in Related Work as an example of architecture-based continual learning and is not load-bearing for GRID's design. The Limitations section appropriately acknowledges that decoding uses task output spaces and may not extend to fully open-ended generation. No equation in the paper reduces BWT or accuracy to a fit or to a self-citation chain, so the circularity burden is not met.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests mainly on benchmark measurements rather than a derivation. The free parameters are standard prompt-tuning hyperparameters plus the compression threshold alpha, whose exact value is not reported. The most consequential assumption is that label sets are known or inferable at inference time, which is also the main limitation.

free parameters (5)
  • alpha (prompt selection threshold) = not stated in main text; sensitivity shown for 0.0 to 1.6 in Table 13
    Controls how many prompts are retained versus merged. The best BWT/memory trade-off is at alpha 0.6 to 0.8, but the exact value used in the main runs is not reported, so it may have been selected on the benchmark.
  • k (representative samples per class) = 1000
    Fixed budget for clustering-based sampling. Table 15 shows performance saturates near 1000 samples per class, so the value was chosen based on the benchmark.
  • Number of clusters C in K-Means = not reported
    Algorithm 2 selects the top k/C samples per cluster, but C is never given in the configuration table. Without C the sampling procedure is under-specified.
  • Prompt length = 10
    Set after a trade-off study in Table 16; longer prompts help BWT but hurt accuracy, so 10 was chosen as a balance.
  • tau (FTC threshold multiplier) = 0.4
    Used to define Forgotten Task Count in Appendix C.2, applied relative to the minimum standalone accuracy across tasks, which yields a low absolute bar for declaring a task forgotten.
assumptions (4)
  • domain assumption Task boundaries are available during training
    Stated in the Limitations section: the method trains one prompt per task and cannot handle fully task-free or streaming settings.
  • domain assumption The output space (label set) of each task is known or inferable at inference for label identification and constrained decoding
    The method needs candidate label sets for the LLM-based label identification and the decoding mask, which limits applicability to open-ended generation, as acknowledged in Limitations.
  • ad hoc to paper Gradient norm on new task data is a valid proxy for prompt redundancy
    The paper supports this with a post-hoc observation of high cosine similarity among low-gradient prompts, not with a derivation.
  • domain assumption A frozen pretrained LLM can perform all tasks when given the right prompts
    Standard for prompt tuning methods; all results depend on the base model's existing capabilities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GRID: Scaling Task-Agnostic Inference in Continual Prompt Tuning." pith.science (2026). https://pith.science/paper/7L6WMI2H

@misc{pith2026250714725,
  author       = {Pith},
  title        = {Pith review of: GRID: Scaling Task-Agnostic Inference in Continual Prompt Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7L6WMI2H}},
  note         = {Machine review of arXiv:2507.14725}
}
read the original abstract

Prompt-based continual learning (CL) offers a parameter-efficient way to adapt large language models (LLMs) across task sequences. However, existing methods often rely on task-aware inference and maintain an expanding set of task-specific prompts, leading to (1) severe performance degradation on earlier tasks when task identifiers are unavailable for prompt selection at inference time, and (2) limited scalability as task sequence grows. We propose GRID, a unified framework designed to address these challenges. GRID incorporates an output-space-aware decoding mechanism that enhances backward transfer by leveraging representative inputs and automatic label semantic normalization, alongside a gradient-guided prompt selection strategy that compresses less informative prompts into a single aggregated representation for scalable, memory-efficient continual learning. Extensive experiments on long-sequence and negative-transfer benchmarks show that GRID improves backward transfer, achieves competitive forward transfer, and substantially reduces prompt memory across encoder-decoder and decoder-only architectures, including T5, Qwen, and LLaMA. Source code is available at https://github.com/AnushkaTi/GRID.

Figures

Figures reproduced from arXiv: 2507.14725 by the authors.

Figure 1
Figure 1. Illustration of GRID, task-agnostic inference in prompt-based continual learning. Multiple tasks with distinct input formats and label spaces are processed using a shared, fixed-capacity prompt pool and a frozen pre-trained model, without access to task identities at inference time. Building on this, continual prompt tuning (CPT) extends PT to the CL setting, enabling sequential task adaptation without modifying the… view at source ↗
Figure 2
Figure 2. Overview of GRID. Given a continual task sequence, GRID first selects representative samples for each task, then trains a task-specific soft prompt while keeping previously learned prompts frozen. After each task, gradient-guided prompt compression scores the frozen prompt queue and retains or merges prompts based on their contribution to the current task. At inference time, GRID uses the compressed prompt queue wit… view at source ↗
Figure 3
Figure 3. Prompt pool compression trajectory under GRID for Order L1. Each leaf represents a task-specific prompt, with colors denoting task categories. Branches indicate prompt consolidation events during compression, while solo branches denote prompts retained independently. The trajectory shows that GRID forms interpretable compressed groups across re￾lated task types while preserving isolated prompts when merg￾ing is less… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Heatmaps of backward transfer scores on previous tasks for Order L1. (A) Progressive Prompts, (B) SHLPT, (C) GRID, and differences (D) C–A, (E) C–B. suffer greater forgetting, whereas smaller models retain prior knowledge better, likely due to less aggressive adaptatio…
Figure 5
Figure 5. Figure 5: Effect of model scaling across T5-Large (0.77B), T5-XL (3B), Qwen-3 (4B), and LLaMA-2 (7B). Results on larger models ( [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Task-level comparison between GRID and per￾task LoRA SFT on MTCL-Bench. Each point reports task￾level accuracy for Qwen and LLaMA, with shaded regions separating math and tool-use tasks. Per-task LoRA SFT trains an independent LoRA adapter for each task and serves as a…
Figure 7
Figure 7. Figure 7: shows that moderate queue sizes (e.g., K=5) provide the best trade-off between accu￾racy, forward transfer, and backward transfer, while larger queues degrade performance due to redun￾dant and weakly task-relevant prompts. Average Accuracy Forward Transfer Backward Tra…
Figure 8
Figure 8. Figure 8: Per-task BWT comparison between our method (blue) and the ProgPrompt (red). Positive bars indicate im￾proved retention of prior tasks. Our method shows consistent BWT gains across task orders, demonstrating its effectiveness in mitigating forgetting across diverse task…
Figure 9
Figure 9. Figure 9: Heatmaps of backward transfer scores on previous tasks for Order L2. Brighter values indicate better retention of earlier tasks. -0.45 -0.45 -0.38 -0.33 -0.64 -0.54 -0.91 -0.52 -0.32 -0.39 -0.34 0.00 -0.20 -0.00 0.00 -0.63 -0.94 -0.57 -0.53 -0.88 -0.71 -0.98 -0.85 -0.8…
Figure 12
Figure 12. Figure 12: Heatmaps of backward transfer scores on previous tasks for Order L4. Brighter values indicate better retention of earlier tasks. multirc wic boolq copa qqp cb rte mnli dbpedia_14 yahoo ag_news amazon yelp imdb sst2 multirc wic boolq copa qqp cb rte mnli dbpedia_14 yah…
Figure 13
Figure 13. Figure 13: Heatmaps of backward transfer scores on previous tasks for Order L5. Brighter values indicate better retention of earlier tasks. sst2 copa ag_news imdb mnli yahoo rte yelp qqp cb amazon dbpedia_14 boolq wic multirc sst2 copa ag_news imdb mnli yahoo rte yelp qqp cb ama…
Figure 17
Figure 17. Figure 17: (a) Left - Backward transfer dynamics across se￾quence (L1). (b) Right - Per-task BWT comparison for Order L1, where GRID (in blue) consistently improves retention over ProgPrompt (in red). 22 [PITH_FULL_IMAGE:figures/full_fig_p022_17.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 27 canonical work pages

  1. [1]

    Online continual learning with maximally interfered retrieval, 2019 a

    Rahaf Aljundi, Lucas Caccia, Eugene Belilovsky, Massimo Caccia, Min Lin, Laurent Charlin, and Tinne Tuytelaars. Online continual learning with maximally interfered retrieval, 2019 a . URL https://arxiv.org/abs/1908.04742

    arXiv 2019

  2. [2]

    Gradient based sample selection for online continual learning

    Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. Advances in neural information processing systems, 32, 2019 b

    work page 2019

  3. [3]

    Rainbow memory: Continual learning with a memory of diverse samples

    Jihwan Bang, Heesu Kim, YoungJoon Yoo, Jung-Woo Ha, and Jonghyun Choi. Rainbow memory: Continual learning with a memory of diverse samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 8218--8227, 2021

    work page 2021

  4. [4]

    Selecting representative data sets

    Tomas Borovicka, Marcel Jirina Jr, Pavel Kordik, and Marcel Jirina. Selecting representative data sets. Advances in data mining knowledge discovery and applications, 12: 0 43--70, 2012

    work page 2012

  5. [5]

    Dokania, Thalaiyasingam Ajanthan, and Philip H

    Arslan Chaudhry, Puneet K. Dokania, Thalaiyasingam Ajanthan, and Philip H. S. Torr. Riemannian Walk for Incremental Learning: Understanding Forgetting and Intransigence, pp.\ 556–572. Springer International Publishing, 2018 a . ISBN 9783030012526. doi:10.1007/978-3-030-01252-6_33. URL http://dx.doi.org/10.1007/978-3-030-01252-6_33

    work page doi:10.1007/978-3-030-01252-6_33 2018

  6. [6]

    Efficient lifelong learning with a-gem

    Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a-gem. arXiv preprint arXiv:1812.00420, 2018 b

    arXiv 2018

  7. [7]

    Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean,...

    arXiv 2022

  8. [8]

    Episodic memory in lifelong language learning

    Cyprien de Masson D'Autume, Sebastian Ruder, Lingpeng Kong, and Dani Yogatama. Episodic memory in lifelong language learning. Advances in Neural Information Processing Systems, 32, 2019

    work page 2019

Show all 54 references
  1. [9]

    Parameter-efficient fine-tuning of large-scale pre-trained language models

    Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence, 5 0 (3): 0 220--235, 2023

    2023

  2. [10]

    Dytox: Transformers for continual learning with dynamic token expansion

    Arthur Douillard, Alexandre Ram \'e , Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 9285--9295, 2022

    2022

  3. [11]

    Memory efficient continual learning with transformers

    Beyza Ermis, Giovanni Zappella, Martin Wistuba, Aditya Rawal, and Cedric Archambeau. Memory efficient continual learning with transformers. Advances in Neural Information Processing Systems, 35: 0 10629--10642, 2022

    2022

  4. [12]

    Ppt: Pre-trained prompt tuning for few-shot learning

    Yuxian Gu, Xu Han, Zhiyuan Liu, and Minlie Huang. Ppt: Pre-trained prompt tuning for few-shot learning. arXiv preprint arXiv:2109.04332, 2021

    2021 arXiv

  5. [13]

    Q-tuning: Queue-based prompt tuning for lifelong few-shot language learning

    Yanhui Guo, Shaoyuan Xu, Jinmiao Fu, Jia Liu, Chaosheng Dong, and Bryan Wang. Q-tuning: Queue-based prompt tuning for lifelong few-shot language learning. arXiv preprint arXiv:2404.14607, 2024

    2024 arXiv

  6. [14]

    Visual prompt tuning

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European conference on computer vision, pp.\ 709--727. Springer, 2022

    2022

  7. [15]

    Towards anytime fine-tuning: Continually pre-trained language models with hypernetwork prompt

    Gangwei Jiang, Caigao Jiang, Siqiao Xue, James Y Zhang, Jun Zhou, Defu Lian, and Ying Wei. Towards anytime fine-tuning: Continually pre-trained language models with hypernetwork prompt. arXiv preprint arXiv:2310.13024, 2023

    2023 arXiv

  8. [16]

    Grad-match: Gradient matching based data subset selection for efficient deep model training, 2021

    Krishnateja Killamsetty, Durga Sivasubramanian, Ganesh Ramakrishnan, Abir De, and Rishabh Iyer. Grad-match: Gradient matching based data subset selection for efficient deep model training, 2021. URL https://arxiv.org/abs/2103.00123

    2021 arXiv

  9. [17]

    Overcoming catastrophic forgetting in neural networks

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of scien...

    2017

  10. [18]

    The power of scale for parameter-efficient prompt tuning

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (eds.), Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing,...

    2021 doi

  11. [19]

    Prefix-tuning: Optimizing continuous prompts for generation

    Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021

    2021 arXiv

  12. [20]

    Learning without forgetting

    Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40 0 (12): 0 2935--2947, 2017

    2017

  13. [21]

    Learning word vectors for sentiment analysis

    Andrew Maas, Raymond E Daly, Peter T Pham, Dan Huang, Andrew Y Ng, and Christopher Potts. Learning word vectors for sentiment analysis. In Proceedings of the 49th annual meeting of the association for computational linguistics: Human language technologies, pp.\ 142--150, 2011

    2011

  14. [22]

    Catastrophic interference in connectionist networks: The sequential learning problem

    Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation, volume 24, pp.\ 109--165. Elsevier, 1989

    1989

  15. [23]

    Coresets for data-efficient training of machine learning models, 2020

    Baharan Mirzasoleiman, Jeff Bilmes, and Jure Leskovec. Coresets for data-efficient training of machine learning models, 2020. URL https://arxiv.org/abs/1906.01827

    2020 arXiv

  16. [24]

    Q.: A survey on transfer learning

    SJ—Yang Pan. Q.: A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22 0 (10): 0 1345--1359, 2010

    2010

  17. [25]

    Lfpt5: A unified framework for lifelong few-shot language learning based on prompt tuning of t5

    Chengwei Qin and Shafiq Joty. Lfpt5: A unified framework for lifelong few-shot language learning based on prompt tuning of t5. arXiv preprint arXiv:2110.07298, 2021

    2021 arXiv

  18. [26]

    Exploring the limits of transfer learning with a unified text-to-text transformer

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21 0 (140): 0 1--67, 2020

    2020

  19. [27]

    Progressive prompts: Continual learning for language models

    Anastasia Razdaibiedina, Yuning Mao, Rui Hou, Madian Khabsa, Mike Lewis, and Amjad Almahairi. Progressive prompts: Continual learning for language models. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=UJTgQBc91_

    2023

  20. [28]

    icarl: Incremental classifier and representation learning

    Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pp.\ 2001--2010, 2017

    2001

  21. [29]

    Sentence-bert: Sentence embeddings using siamese bert-networks

    Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 11 2019. URL https://arxiv.org/abs/1908.10084

    2019 arXiv

  22. [30]

    Representation matters: Assessing the importance of subgroup allocations in training data

    Esther Rolf, Theodora T Worledge, Benjamin Recht, and Michael Jordan. Representation matters: Assessing the importance of subgroup allocations in training data. In International Conference on Machine Learning, pp.\ 9040--9051. PMLR, 2021

    2021

  23. [31]

    AdapterDrop : O n the efficiency of adapters in transformers

    Andreas R \"u ckl \'e , Gregor Geigle, Max Glockner, Tilman Beck, Jonas Pfeiffer, Nils Reimers, and Iryna Gurevych. AdapterDrop : O n the efficiency of adapters in transformers. In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (eds.), Proceedings of...

    2021 doi

  24. [32]

    Progressive neural networks

    Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016

    2016 arXiv

  25. [33]

    Shrey Satapara and P. K. Srijith. TL - CL : Task and language incremental continual learning. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (eds.), Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pp.\ 12123--12142, Miami, Florida,...

    2024 doi

  26. [34]

    Continual learning with deep generative replay

    Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay. Advances in neural information processing systems, 30, 2017

    2017

  27. [35]

    Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning

    James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEE...

    2023

  28. [36]

    Lamol: Language modeling for lifelong language learning

    Fan-Keng Sun, Cheng-Hao Ho, and Hung-Yi Lee. Lamol: Language modeling for lifelong language learning. arXiv preprint arXiv:1909.03329, 2019

    1909 arXiv

  29. [37]

    Three scenarios for continual learning

    Gido M Van de Ven and Andreas S Tolias. Three scenarios for continual learning. arXiv preprint arXiv:1904.07734, 2019

    1904 arXiv

  30. [38]

    Efficient continual learning with modular networks and task-driven priors

    Tom Veniat, Ludovic Denoyer, and Marc'Aurelio Ranzato. Efficient continual learning with modular networks and task-driven priors. arXiv preprint arXiv:2012.12631, 2020

    2012 arXiv

  31. [39]

    Glue: A multi-task benchmark and analysis platform for natural language understanding

    Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018

    2018 arXiv

  32. [40]

    Superglue: A stickier benchmark for general-purpose language understanding systems

    Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. Superglue: A stickier benchmark for general-purpose language understanding systems. Advances in neural information processing systems, 32, 2019

    2019

  33. [41]

    Pre-trained language models and their applications

    Haifeng Wang, Jiwei Li, Hua Wu, Eduard Hovy, and Yu Sun. Pre-trained language models and their applications. Engineering, 25: 0 51--65, 2023 a

    2023

  34. [42]

    A comprehensive survey of continual learning: Theory, method and application

    Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  35. [43]

    Multitask prompt tuning enables parameter-efficient transfer learning

    Zhen Wang, Rameswar Panda, Leonid Karlinsky, Rogerio Feris, Huan Sun, and Yoon Kim. Multitask prompt tuning enables parameter-efficient transfer learning. In The Eleventh International Conference on Learning Representations, 2023 b . URL https://openreview.net/forum?id=Nk2pDtuhTq

    2023

  36. [44]

    Dualprompt: Complementary prompting for rehearsal-free continual learning

    Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European conference on computer vision, pp.\ 631--648. Springe...

    2022

  37. [45]

    Learning to prompt for continual learning

    Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 139--149, 2022 b

    2022

  38. [46]

    Efficient meta lifelong-learning with limited memory

    Zirui Wang, Sanket Vaibhav Mehta, Barnabas Poczos, and Jaime Carbonell. Efficient meta lifelong-learning with limited memory. In Bonnie Webber, Trevor Cohn, Yulan He, and Yang Liu (eds.), Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (E...

    2020 doi

  39. [47]

    Mitigate negative transfer with similarity heuristic lifelong prompt tuning

    Chenyuan Wu, Gangwei Jiang, and Defu Lian. Mitigate negative transfer with similarity heuristic lifelong prompt tuning. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar (eds.), Findings of the Association for Computational Linguistics: ACL 2024, pp.\ 10944--10959, Bangkok, Tha...

    2024 doi

  40. [48]

    Parameter-efficient fine-tuning methods for pretrained language models: A critical review and assessment, 2023

    Lingling Xu, Haoran Xie, Si-Zhao Joe Qin, Xiaohui Tao, and Fu Lee Wang. Parameter-efficient fine-tuning methods for pretrained language models: A critical review and assessment, 2023. URL https://arxiv.org/abs/2312.12148

    2023 arXiv

  41. [49]

    C on T in T in: Continual learning from task instructions

    Wenpeng Yin, Jia Li, and Caiming Xiong. C on T in T in: Continual learning from task instructions. In Smaranda Muresan, Preslav Nakov, and Aline Villavicencio (eds.), Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers...

    2022 doi

  42. [50]

    Lifelong learning with dynamically expandable networks

    Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id=Sk7KsfW0-

    2018

  43. [51]

    Continual learning through synaptic intelligence

    Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International conference on machine learning, pp.\ 3987--3995. PMLR, 2017

    2017

  44. [52]

    Character-level convolutional networks for text classification

    Xiang Zhang, Junbo Zhao, and Yann LeCun. Character-level convolutional networks for text classification. Advances in neural information processing systems, 28, 2015

    2015

  45. [53]

    Continual prompt tuning for dialog state tracking

    Qi Zhu, Bing Li, Fei Mi, Xiaoyan Zhu, and Minlie Huang. Continual prompt tuning for dialog state tracking. In Smaranda Muresan, Preslav Nakov, and Aline Villavicencio (eds.), Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Lon...

    2022 doi

  46. [54]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.