Artificial Intelligence · arXiv:2610.15492 v1 · submitted 11 Oct 2026 · 19 references
When Skills Remain But Access Fails: Factorized Skill Memory for Continual Skill Internalization
The paper says
“Skill internalization aims to convert externally specified procedural skills into persistent model capabilities, allowing language-model agents to execute them without retrieving or reinjecting skill descriptions at inference time.”
more of the abstract
Yet when new skills are learned continually, whether previously internalized skills remain usable is still poorly understood. We study forgetting in continual skill internalization and uncover a clear hierarchical structure. Early in subsequent training, a model may lose the ability to invoke an old skill from the raw request, while supplying only the correct canonical skill name substantially restores the learned behavior; with further optimization, this conditional execution also degrades. We call these two stages skill-addressing forgetting and skill-execution forgetting. Guided by this finding, we introduce Factorized Skill Memory (FSM), which separates continual skill memory into a plastic Address Memory and a stable Execution Memory. At each stage, FSM trains a new LoRA executor using only current-stage data, freezes it after acquisition, and appends it to Execution Memory; in parallel, Address Memory incrementally learns to map current requests to executors without replaying historical requests or execution trajectories. At inference time, Address Memory selects one executor from the raw request, and only that frozen executor performs a single language-model generation. On a five-stage benchmark built from 100 TaskBench skills, FSM achieves 65.95% mean exact match and 93.38% addressing accuracy across two task orders, improving over the strongest replay-free parameter-isolation baseline by 14.3 percentage points while remaining within 1.9 points of Replay, which retains historical training examples.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 14 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. The two-stage forgetting hierarchy is real and worth publishing, but don't take the 0 to 69.85% recovery at face value: the address leaks the gold node set, and the paper knows it. This line is wrong
2 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 19 references say.
The paper says, at Section 4.1, Appendix B.1
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.63):
“These controls rule out generic non-target cues, but the correct skill address still reveals target node identity.”
Pith says: Objection 1 of 2 · major
The paper is explicit that the correct skill address also reveals target node identity and that the diagnostic is not an information-neutral access probe. The controls rule out the address alone as the answer generator: an incorrect address gives 6.86% EM, a semantic description gives 0%, and a frozen base model gives 0% EM despite 92.65% correct skill selection.
more of objection 1
What they do not rule out is a name-triggered reconstruction: the post-forgetting checkpoint, which has seen both T2 and T3 training, could use the semantically meaningful canonical name as a strong in-context cue to assemble a plausible TaskBench graph from format competence and world knowledge, without retrieving a retained T2 request-conditioned procedure. Since the 0% to 69.85% recovery is the primary evidence for the 'skills remain but access fails' claim, this residual confound is load-bearing. I request either (i) a control in which skills are associated during acquisition with opaque, non-semantic addresses and the post-forgetting checkpoint is probed with those opaque addresses, or (ii) a revision that limits the claim to address-conditioned execution and explicitly leaves the underlying mechanism open.
The paper says, at Section 4.2, Figure 2
The review did not quote this passage.
Pith says: Objection 2 of 2 · major
The temporal ordering that defines the two regimes, namely closed-book collapse while address-conditioned execution survives and then address-conditioned execution erodes, is measured on a single task transition, T2 to T3, albeit with three training procedures. The bottom matrices of Figure 2 and the correct-skill-address rows of Table 2 are final-checkpoint diagnostics; they establish static recoverability but not the ordering of addressing forgetting before execution forgetting.
more of objection 2
Given that the title and conclusion generalize to continual skill internalization, please either report the trajectory for at least one additional transition or explicitly limit the general claim to the examined transition.
5 minor objections
The paper says, at Section 3 vs. Appendix C.4, Table A13
Pith says: minor objection 1 of 5
The first request-only zero for SKILL0 is reported as step 10 in Section 3 but as step 18 in Table A13; these values should be aligned or the difference explained.
The paper says, at Section 6.1, Table 2
Pith says: minor objection 2 of 5
The stated improvement of 14.3 percentage points over Progressive-LoRA does not match the reported macro averages (65.95 vs. 51.98 is 13.97); please define whether the comparison is a mean of per-order differences or a difference of macro averages.
show 3 more
The paper says, at Appendix A.2
Pith says: minor objection 3 of 5
Please state explicitly how unparseable or malformed model outputs are handled in the EM computation (counted as failures, excluded, or repaired) and report the number of such outputs per condition, since every headline number is an exact-match percentage.
The paper says, at Section 5.2, Appendix A.5
Pith says: minor objection 4 of 5
The Address Memory implementation is under-specified: please provide the hashing dimensionality, tokenization or n-gram settings, Naive Bayes update order, and smoothing parameters so that the replay-free address learner can be reproduced exactly.
The paper says, at Section 7
Pith says: minor objection 5 of 5
The sentence 'FSM reaches 65.9593.38percentage points' is missing the intervening text; it should read something like 'FSM reaches 65.95% mean EM and 93.38% routing accuracy'.
What would settle it
Retrain the T2 stage with canonical skill identifiers replaced by opaque random tokens (e.g., SKILL_7F3A) in both the training targets and the reactivation prompts, keeping requests, graphs, LoRA recipe, and the T2→T3 transition identical; then rerun the closed-book and correct-address evaluation at the post-forgetting checkpoint. If addressed EM remains near 69.85% after closed-book collapse, the recovery does not depend on the semantic content of the canonical name and the execution-memory interpretation is supported. If addressed EM drops toward the incorrect-address control (6.86%), the original recovery is substantially a semantic-cue artifact of the gold node name, and the 'skills remain but access fails' claim must be weakened to 'name-conditioned reconstruction remains.'
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 14 Oct 2026
- Reader holds up, with conditions
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 2 major objections, 5 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
The paper earns a serious referee. The strongest thing in it is not FSM—that is a known per-stage LoRA freezing plus router family, and the authors say so—but the empirical decomposition it documents: after continual training moves off an old skill family, closed-book EM on 408 T2 requests is 0, while supplying the correct skill name lifts full-task EM to 69.85%; further T3 optimization degrades even that conditional recovery. Plenty of papers assert access-vs-memory separation; this one actually measures the two stages over time and gets the ordering right across SFT, GRPO, and SKILL0.
Credit where it is due. The diagnostic includes controls—wrong address 6.86%, semantic description 0%, frozen base 0% despite 92.65% name recognition—and the authors explicitly state that the correct address reveals target node identity and is a diagnostic, not an information-neutral probe. That is honest framing. Two task orders, two model scales, an oracle-versus-learned-routing decomposition, and a substantial appendix make this a more complete package than the average arXiv method paper.
The soft spots are real but proportionate. The stress-test concern is the main one: the correct skill address is the benchmark's exact procedure identifier, and for multi-node requests that is effectively the gold node set. It removes skill selection, composition, and part of the request-to-execution mapping. The 0-to-69.85% number therefore overstates how much request-conditioned procedural knowledge is retained; what is shown is that execution under a supplied identity is substantially recoverable. The controls rule out the address alone as an answer generator, but they do not separate a retained T2 procedure from a name-triggered reconstruction built from TaskBench format competence plus world knowledge. The paper's caveat is honest, but this claim is load-bearing for the slogan “skills remain but access fails.”
Other issues are evaluation hygiene: no standard deviations or seed analysis, no code or data release, Replay is a single configuration, and the paper does not state how parse failures are handled. All fixable. I would push for a control using opaque per-skill identifiers assigned during training: if an arbitrary trained token recovers execution, the procedural-retention reading gets much stronger; if only natural-language canonical names work, the reconstruction reading stays live.
Bottom line: conditional acceptance, with code, multi-seed error bars, and that opaque-address control. The hierarchy result deserves refereeing, not desk rejection.
The strongest objection
Section 4.1's reactivation diagnostic supplies the benchmark's exact procedure identifier(s) for each request, and for multi-node requests this is effectively the gold node set itself. The paper acknowledges this in Appendix B.1 ('the correct skill address still reveals target node identity') and runs controls: an incorrect address gives 6.86% EM, a semantic description 0%, and a frozen base model 0% EM despite 92.65% correct skill selection. Those controls rule out the address alone as an answer generator and rule out generic non-target cues. What they do not rule out is that the post-forgetting checkpoint, which has seen both T2 and T3 training, uses the semantically meaningful canonical name as a trigger to reconstruct a plausible TaskBench graph from general format competence plus world knowledge, rather than retrieving a retained T2 request-conditioned procedure. Because the address also removes the skill-selection and composition subproblem, the measured 69.85% EM is conditional execution with the hardest part of the skill supplied externally. The central hierarchy claim—'skills remain but access fails'—depends on interpreting the 0→69.85% gap as access loss, so this confound is load-bearing even though the authors are explicit that the address is a diagnostic rather than an information-neutral probe.
The central claim is the hierarchical forgetting structure, and its load-bearing premise is that the 0→69.85% recovery under the correct skill address reflects retained execution after request-side access fails. Section 4.1 and Appendix B.1 explicitly acknowledge that the address reveals target node identity and supply three controls: incorrect address (6.86%), semantic description (0%), and frozen base (0% EM with 92.65% selection). These controls rule out the address alone as an answer generator and rule out generic non-target cues. What remains unresolved is whether the recovered EM is produced by a retained T2 request-conditioned procedure or by a general TaskBench-format competence triggered by the semantically meaningful canonical name; TaskBench identifiers such as daily_bill_payment carry real semantics, so the address may be doing part of the 'which skill and how' work. The frozen-base control is on an untrained model and does not isolate this specific confound. I therefore regard the concern as load-bearing but not fatal: FSM's headline results (65.95% mean EM, 93.38% routing accuracy) do not depend on the address diagnostic being information-neutral, and the existing controls substantially narrow the alternative explanation. The reader's weakest_assumption identifies the same point; I agree partially because the authors have already constrained it, but the opaque-identifier retraining test would settle whether semantic content of the address is essential. The absence of seeds, error bars, and code is a separate reproducibility concern that independently supports keeping the verdict CONDITIONAL rather than ACCEPT.
The referee's report
The contribution is useful if the central interpretation holds: an explicit behavioral distinction between access to a skill and execution of that skill, a simple modular method that operationalizes the distinction, and a careful empirical study with multiple controls. Strengths of the paper include the incorrect-address, semantic-description, and frozen-base controls; two task orders; a second model scale; detailed information-budget accounting; and transparent acknowledgement that the address diagnostic is not information-neutral. The main scientific claim, however, depends on interpreting the 0% to 69.85% address-conditioned recovery as evidence that execution remains while request-to-skill access has failed. Because the correct address reveals the target node identity, that interpretation needs either an additional control or a more modest framing. The FSM results themselves are separable from this diagnostic issue and are reported with sufficient detail to be reproduced.
Pith's reading of the paper (a machine's summary, not the authors')
The paper studies whether skills a language-model agent has internalized remain usable when new skills are learned afterward. Its central finding is that forgetting is not a single collapse but a two-stage hierarchy: early in subsequent training, the model loses the ability to invoke an old skill from the raw request alone, while supplying the canonical skill name restores most of the learned behavior; with continued optimization, even that name-assisted execution degrades. The paper calls these skill-addressing forgetting and skill-execution forgetting, and reports that at a checkpoint failing all 408 old requests, the correct address lifts exact-match success from 0% to 69.85%. Guided by this diagnosis, it introduces Factorized Skill Memory, which freezes each stage's skill executor and keeps only a lightweight request-to-shard index plastic, reaching 65.95% mean exact match without replaying historical examples.
What carries the argument. The load-bearing object is the correct-skill-address reactivation diagnostic: take a post-forgetting checkpoint that scores zero closed-book, supply only the benchmark's canonical skill identifier, and measure whether execution returns. This diagnostic splits forgetting into two regimes — skill-addressing forgetting, where request-side access is lost but name-assisted execution stays intact, and skill-execution forgetting, where even name-assisted execution erodes. Factorized Skill Memory operationalizes the split at coarser shard granularity: each stage trains a fresh rank-16 LoRA executor on the frozen base, freezes it into Execution Memory, and lets an incremental Naive Bayes Address Memory learn request-to-shard identity from only current-stage text, with a single hard memory read at inference routing to one frozen shard that still performs within-shard skill selection and composition.
Core claim. On its own terms, the paper claims to establish a hierarchical forgetting structure in continual skill internalization and a method that follows it. The structure is that closed-book access to an old skill fails before the skill's execution is lost: a correct-skill-address reactivation diagnostic shows a checkpoint at 0% full-task EM recovering to 69.85% when the canonical skill name is supplied, and only later does continued optimization erode execution under the correct address, from 84.56% to 57.60% in GRPO. Because the address also reveals the gold node identity, the paper uses it as a diagnostic rather than a pure access probe and runs controls (incorrect address 6.86%, semantic description 0%, frozen base 0% EM) that localize the recovery to learned execution rather than address recognition. Guided by this finding, Factorized Skill Memory separates the two problems: per-stage LoRA executors are frozen into an append-only Execution Memory while a plastic Address Memory learns request-to-shard associations without replaying historical data, reaching 65.95% mean exact match and 93.38% routing accuracy, within 1.9 points of Replay without retaining training examples.
The premise it rests on. The load-bearing premise is that the recovery from 0% to 69.85% under the correct skill address reflects restored access to intact procedural knowledge, not the address itself acting as a strong in-context cue that reveals the target node and thereby leaks behavior the model could reconstruct without any retained skill.
If the paper is right (Pith's inferences, not the authors')
Forgetting in continual skill internalization is not a single collapse but a two-stage hierarchy, so retention diagnostics must separate request-side access from execution rather than reporting one closed-book EM number.
Freezing acquired execution and keeping only a plastic address index recovers most Replay-style retention without storing any historical requests, at 65.95% mean EM versus 69.32% for Replay.
Error is decomposable: T2 is mostly access-limited and T1 mostly execution-limited, so the correct remedy differs by skill family — better routing versus better executors.
The same addressed-reactivation ordering appears across SFT, GRPO, and SKILL0, suggesting the hierarchy is a property of continual optimization generally rather than of one training recipe.
The addressing/execution split likely generalizes beyond skill internalization to any procedural capability in language models, such as tool-use policies or retrieval-augmented workflows, where "the name unlocks the procedure" could mask erasure in standard evaluation.
A genuinely information-neutral probe — a request paraphrase or a hash that reveals no node identity — would separate the diagnostic interpretation from label leakage, since the paper's own controls show the canonical name leaks the target node identity.
The learning-rate sweep's abrupt five-step transition (T2 from 79.41 to 0 while T3 jumps to 18.98) suggests execution forgetting stems from an overtraining switch rather than gradual decay, pointing to an adaptive or mixed-initiative rate schedule as a testable extension.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Skill internalization aims to convert externally specified procedural skills into persistent model capabilities, allowing language-model agents to execute them without retrieving or reinjecting skill descriptions at inference time. Yet when new skills are learned continually, whether previously internalized skills remain usable is still poorly understood. We study forgetting in continual skill internalization and uncover a clear hierarchical structure. Early in subsequent training, a model may lose the ability to invoke an old skill from the raw request, while supplying only the correct canonical skill name substantially restores the learned behavior; with further optimization, this conditional execution also degrades. We call these two stages skill-addressing forgetting and skill-execution forgetting. Guided by this finding, we introduce Factorized Skill Memory (FSM), which separates continual skill memory into a plastic Address Memory and a stable Execution Memory. At each stage, FSM trains a new LoRA executor using only current-stage data, freezes it after acquisition, and appends it to Execution Memory; in parallel, Address Memory incrementally learns to map current requests to executors without replaying historical requests or execution trajectories. At inference time, Address Memory selects one executor from the raw request, and only that frozen executor performs a single language-model generation. On a five-stage benchmark built from 100 TaskBench skills, FSM achieves 65.95% mean exact match and 93.38% addressing accuracy across two task orders, improving over the strongest replay-free parameter-isolation baseline by 14.3 percentage points while remaining within 1.9 points of Replay, which retains historical training examples.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-14
The authors' main claim, as the classifier picked it out 0.57
“On a five-stage benchmark built from 100 TaskBench skills, FSM achieves 65.95% mean exact match and 93.38% addressing accuracy across two task orders, improving over the strongest replay-free parameter-isolation baseline by 14.3 percentage points while remaining within 1.9 points of Replay, which retains historical training examples.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.34 · human participants 0.03 · controlled comparison 0.96 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Skill internalization aims to convert externally specified procedural skills into persistent model capabilities, allowing language-model agents to execute them without retrieving or reinjecting skill descriptions at inference time. | background 1.00 | 0.06 | 0.03 | asserted 0.99 |
| Yet when new skills are learned continually, whether previously internalized skills remain usable is still poorly understood. | gap 1.00 | 0.05 | 0.02 | asserted 0.95 |
| We study forgetting in continual skill internalization and uncover a clear hierarchical structure. | contribution 0.73 | 0.96 | 0.15 | asserted 1.00 |
| Early in subsequent training, a model may lose the ability to invoke an old skill from the raw request, while supplying only the correct canonical skill name substantially restores the learned behavior; with further optimization, this conditional execution also degrades. | result 1.00 | 0.88 | 0.27 | hedged possibility 0.82 |
| We call these two stages skill-addressing forgetting and skill-execution forgetting. | interpretation 0.50 | 0.84 | 0.11 | asserted 0.97 |
| Guided by this finding, we introduce Factorized Skill Memory (FSM), which separates continual skill memory into a plastic Address Memory and a stable Execution Memory. | contribution 0.72 | 0.88 | 0.18 | asserted 1.00 |
| At each stage, FSM trains a new LoRA executor using only current-stage data, freezes it after acquisition, and appends it to Execution Memory; in parallel, Address Memory incrementally learns to map current requests to executors without replaying historical requests or execution trajectories. | method 1.00 | 0.90 | 0.10 | asserted 0.99 |
| At inference time, Address Memory selects one executor from the raw request, and only that frozen executor performs a single language-model generation. | method 1.00 | 0.71 | 0.08 | asserted 1.00 |
| On a five-stage benchmark built from 100 TaskBench skills, FSM achieves 65.95% mean exact match and 93.38% addressing accuracy across two task orders, improving over the strongest replay-free parameter-isolation baseline by 14.3 percentage points while remaining within 1.9 points of Replay, which retains historical training examples. | result 1.00 | 0.96 | 0.40 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-14: 1 request, 7,095 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (4)
Figures are the authors', reproduced from the paper.
References (19)
- [2] Zhiyuan Jiang, Fangrui Huang, Hanwen Xing, Xander Wu, Yipeng Gao, Rui Cao, Mengdi Wang, Shilong Liu, and Yijiang Li. Demystifying agent skills: Why they work—until they don’t.arXiv preprint arXiv:2608.14036, on PithPith's reviewarXiv
- [5] In the same protocol, five prompt-order variants reached 18 Table A10: Current self-confidence Graph EM by order prefix (percent). Dashes mark task families that had not entered the stream. Order Prefix T1 T2 T3 T4 T5 O1 n= 145.02 – – – – n= 242.82 82.60 – – – n= 342.36 72.06 ... on Pith
- [6] Zhengxi Lu, Zhiyuan Yao, Jinyang Wu, Chengcheng Han, Qi Gu, Xunliang Cai, Weiming Lu, Jun Xiao, Yueting Zhuang, and Yongliang Shen. Skill0: In-context agentic reinforcement learning for skill internalization.arXiv preprint arXiv:2604.02268, on PithPith's reviewarXiv
- [11] Gido M. van de Ven and Andreas S. Tolias. Three scenarios for continual learning.arXiv preprint arXiv:1904.07734, on PithPith's reviewarXiv
- [12] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. V oyager: An open-ended embodied agent with large language models.arXiv preprint arXiv:2305.16291, 2023a. Mingyang Wang, Heike Adel, Lukas Lange, Jannik Str ¨otgen, a... on PithPith's reviewarXiv
- [13] doi: 10.18653/v1/2024.naacl-short.39. URL https://aclanthology.org/2024.naacl-short.39/. Xiao Wang, Tianze Chen, Qiming Ge, Han Xia, Rong Bao, Rui Zheng, Qi Zhang, Tao Gui, and Xuanjing Huang. Orthogonal subspace learning for language model continual learning. InFindings of th... on PithPith's reviewarXiv
- [14] doi: 10.20944/preprints202605. 1276.v1. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. InInternational Conference on Learning Representations, on Pithdoi:10.20944/preprints202605
- [15] Aofan Yu, Chenyu Zhou, Tianyi Xu, Zihan Guo, Rong Shan, Zhihui Fu, Jun Wang, Weiwen Liu, Yong Yu, Weinan Zhang, and Jianghao Lin. Latentskill: From in-context textual skills to in-weight latent skills for llm agents.arXiv preprint arXiv:2606.06087, on PithPith's reviewarXiv
- [16] Tianyi Zhang and Zhonghao Qi. Skill-to-lora: From using skills to learning behaviors for token- efficient llm agents.arXiv preprint arXiv:2606.16769, on PitharXiv
- [17] Xuan Zhao, Haonan He, Qingyu Yang, Minglei Li, Jingqi Ye, Zelin Tan, Bo Wan, and Peng Ye. Parametric skills.arXiv preprint arXiv:2606.30015, on PithPith's reviewarXiv
- [18] 12 A EXPERIMENTAL DETAILS AND REPRODUCIBILITY This appendix records the benchmark protocol, evaluation contract, metrics, FSM implementation, information budgets, and safeguards used for the main results. A.1 CONTINUAL INTERNALIZATION PROTOCOL At stage t, the learner receives ... on Pith
- [2017] Xiangyi Li et al. Skillsbench: Benchmarking how well agent skills work across diverse tasks.arXiv preprint arXiv:2602.12670, on PithPith's reviewarXiv
- [2018] Hongxiang Lin, Zhirui Kuai, Erpeng Xue, and Lei Wang. Skillc: Learning autonomous skill internalization in llm agents via contrastive credit assignment.arXiv preprint arXiv:2605.27899, on PithPith's reviewarXiv
- [2019] doi: 10.1016/j.neunet.2019.01.012. Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual adapters. InAdvances in Neural Information Processing Systems, volume 30, 2017a. Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg S... on PithPith's reviewarXiv
- [2021] doi: 10.18653/v1/2021.emnlp-main.590. 10 Bo Mao, Jie Zhou, Yutao Yang, Xin Li, Xian Wei, Qin Chen, Xingjiao Wu, and Liang He. Learning while acting: A skill-enhanced test-time co-evolution framework for online lifelong learning agents. arXiv preprint arXiv:2606.04815, on PithPith's reviewarXiv
- [2022] doi: 10.1109/TPAMI.2021.3057446. Yuchen Guo, Junli Gong, Weicheng Wang, Hongmin Cai, Yiu-ming Cheung, and Weifeng Su. Peam: Parametric embodied agent memory through contrastive internalization of experience in minecraft. arXiv preprint arXiv:2605.27762, on PithPith's reviewarXiv
- [2023] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y . K. Li, Y . Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, on PithPith's reviewarXiv
- [2024] doi: 10.1162/tacl a 00638. David Lopez-Paz and Marc’Aurelio Ranzato. Gradient episodic memory for continual learning. In Advances in Neural Information Processing Systems, volume 30, on PithPith's reviewdoi:10.1162/tacl
- [2026] Andrea Madotto, Zhaojiang Lin, Zhenpeng Zhou, Seungwhan Moon, Paul Crook, Bing Liu, Zhou Yu, Eunjoon Cho, Pascale Fung, and Zhiguang Wang. Continual learning in task-oriented dialogue systems. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Proces... on Pith
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment