Pith. sign in

REVIEW 3 major objections 7 minor 1 cited by

Language models need a periodic sleep stage that expands capacity, distills short-term memories into long-term parameters, and self-improves by generating synthetic dreams.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-13 07:43 UTC pith:R3S7Z76P

load-bearing objection Solid empirical package for sleep-style consolidation, but almost every strong result rides on the Hope/CMS multi-frequency backbone rather than a clean Sleep isolation. the 3 major comments →

arxiv 2606.03979 v2 pith:R3S7Z76P submitted 2026-06-02 cs.LG cs.AI

Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories

Ali Behrouz , Farnoosh Hashemi , Adel Javanmard , Vahab Mirrokni This is my paper
classification cs.LG cs.AI
keywords continual learningmemory consolidationknowledge distillationin-context learningself-improvementlarge language modelscatastrophic forgettingsynthetic data
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Current large language models learn well in the moment through in-context learning but cannot turn that fragile, temporary knowledge into lasting parametric skill without expensive retraining or catastrophic forgetting. This paper argues that a continual learner has no fixed train/test split; instead it should alternate wake periods of taking in new data with sleep periods that consolidate and improve. Sleep has two stages: Knowledge Seeding, which unlocks new low-rank experts in slower memory blocks and distills a smaller version of the model upward into them, and Dreaming, which uses reinforcement learning to generate and select synthetic data that rehearses new knowledge. Experiments on continual classification, novel-language translation, long-context reasoning, fact incorporation, and few-shot abstraction show gains over pure in-context learning, regularization methods, and prior self-adaptation techniques. If the paradigm works, models could keep learning throughout their lifetime without wiping prior capabilities.

Core claim

A periodic Sleep stage—first expanding lower-frequency parameters and distilling a smaller-self into them via generalized on-policy distillation plus RL imitation, then generating a curriculum of synthetic dreams—enables continual transfer of fragile in-context knowledge into stable long-term parameters while supporting recursive self-improvement without human supervision.

What carries the argument

Knowledge Seeding: an upward self-distillation in which a smaller, higher-frequency version of the model (teacher) is distilled into newly expanded low-rank experts inside a larger, lower-frequency student, combining on-policy token-level divergence with an RL imitation reward that aligns both semantic and token-level continuations.

Load-bearing premise

That gradually adding and then freezing low-rank experts, while resetting the high-frequency adapters after each consolidation, is enough to keep old knowledge intact across many sleep cycles without unbounded growth or eventual interference.

What would settle it

Run a long multi-cycle continual stream (dozens of sequential tasks or languages) with the full Sleep schedule and measure whether accuracy on the earliest tasks stays within a few points of the single-task baseline; a sharp drop after many consolidations would falsify the capacity-and-freeze premise.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper proposes a Sleep paradigm for continual LLM learning, replacing the train/test split with wake and sleep phases. Sleep has two stages: (1) Memory Consolidation via Knowledge Seeding—an upward distillation that expands lower-frequency sparse MoE experts with low-rank adapters and transfers knowledge from a smaller-self (higher-frequency) state using Generalized Knowledge Distillation (on-policy mix) plus an RL Learning-to-Imitate objective with semantic and Levenshtein rewards; (2) Dreaming—an RL self-improvement loop that generates synthetic dreams, scores them by SFT gradient importance, injects random-expert novelty, and updates via ReSTEM/SEAL-style rewards. Built on Nested Learning Continuum Memory Systems (multi-frequency MLP blocks), the method is evaluated on class-incremental classification (CLINC, Banking, DBpedia), long-context QA/retrieval, continual novel-language translation, BABILong, SQuAD knowledge incorporation, few-shot ARC, and math reasoning (AIME/HMMT), with ablations on imitation learning, expansion, dreaming, and selection.

Significance. If the gains are attributable to Sleep rather than the Nested Learning substrate, the work would be a meaningful contribution to lifelong LLM adaptation: it gives a concrete offline mechanism for moving fragile in-context knowledge into stable parameters, pairs capacity expansion with freeze-and-reset, and combines on-policy distillation with imitation learning in a way that is distinct from concurrent fixed-capacity OPSD recipes. Strengths include multi-task external benchmarks, component ablations that generally move in the claimed direction (Tables 1–4, Figures 3–6), and explicit positioning against SEAL and later OPSD literature. The wake/sleep lifecycle framing is useful even if some design choices remain heuristic.

major comments (3)
  1. [§4 Empirical Results; Table 3; Figures 3–6] Central claim isolation (§4, Figs. 3–6, Table 3): Almost all positive continual, long-context, and BABILong results are reported on Hope/CMS multi-frequency backbones with nested update frequencies and sparse MoE experts (§2.2, §3.2–3.3). The pure-Transformer rows in Table 3 (knowledge incorporation) improve modestly over SEAL (48.1 vs 46.7 single-passage) and remain weaker than the four-level variant. Without a controlled ablation that applies the same Knowledge Seeding + Dreaming schedule, expansion/freeze/reset recipe, and multi-task sequences to a standard frozen Transformer or LoRA-only backbone, it is not possible to attribute gains to Sleep as stated for general “Language Models” rather than to the Nested Learning continuum-memory substrate on which Sleep is defined.
  2. [§3.2 Parameter Expansion; §3.3 Knowledge Seeding] Multi-cycle capacity and interference (§3.2–3.3): Knowledge Seeding repeatedly adds low-rank experts to lower-frequency blocks, freezes old parameters, consolidates, then resets high-frequency low-rank adapters. The premise that this prevents interference across many sleep cycles without unbounded growth or eventual collapse is load-bearing for continual learning, yet experiments use short task sequences (e.g., two languages in CTNL; class-incremental splits) without long multi-cycle stress tests that track parameter growth, retention of early tasks after many consolidations, or collapse under repeated Dreaming. A multi-cycle retention curve or capacity budget analysis is needed to support the claim that expansion + freeze + reset is sufficient.
  3. [§3.4 Dreaming; Table 3; Table 1] Dreaming vs. consolidation contribution (Table 3, Table 1 ablations): Removing Dreaming drops SQuAD accuracy sharply (35.7/36.2 vs 48.1/44.3), while consolidation-only ablations on math (Table 1/2) still beat SFT/GRPO. The paper asserts that the two-step design is more robust to CF than iterative self-improvement alone, but does not report a head-to-head multi-task continual sequence where Dreaming is applied without prior Knowledge Seeding (or vice versa) under matched compute. That comparison is needed to justify the NREM/REM-style ordering as more than a convenient stack of SEAL-style self-edits on top of CMS.
minor comments (7)
  1. [§2.1 Notation] Notation: θ vs 𝜽 and frequency superscripts are dense; a short glossary of C^(ℓ), f_ℓ, SKS, LTI, GKD would help §2–3.
  2. [§3.3 Learning to Imitate] Eq. (3)–(4): semantic reward is described as a frozen RM giving 0/1, but the RM is not named or validated; absolute reward uses Levenshtein with threshold z0—state default z0 and sensitivity.
  3. [Figure 1] Figure 1 caption discusses Hope architecture but the main text introduces CMS first; align figure labels with §2.2 terminology.
  4. [Appendix A.4] Appendix A.4 on concurrent OPSD is thorough but long; a short main-text paragraph summarizing the four positioning axes would improve accessibility.
  5. [Tables 1–2] Table 1 vs Table 2: both titled for math reasoning with overlapping Qwen3-8B rows; clarify which is ablation-only vs full comparison.
  6. [Appendix B.5] Efficiency note (B.5): SFT is 4× cheaper per step; the wall-clock-to-target claim is useful—report absolute GPU-hours for Sleep vs SFT/GRPO on AIME.
  7. [Abstract; §1] Typos/style: “asmaller-self”, “alargernetwork”, “upwarddistillation”, and similar missing spaces appear in the abstract and §1; clean for camera-ready.

Circularity Check

0 steps flagged

No derivation-by-construction circularity; empirical method paper whose results rest on external benchmarks, with only non-load-bearing self-citation of the Nested Learning substrate.

full rationale

The paper proposes an architectural/training procedure (periodic parameter expansion + Knowledge Seeding via GKD-style on-policy distillation plus RL imitation, followed by Dreaming) and evaluates it on external, publicly defined benchmarks (CLINC, Banking, DBpedia, RULER MK-NIAH, LongHealth, QASPER, MTOB/Manchu CTNL, BABILong, SQuAD, ARC, AIME/HMMT). There is no closed-form prediction, uniqueness theorem, or fitted parameter that is later re-labeled as a derived result; the objectives L_KS and the dreaming reward are optimization targets, not claims that force the reported accuracies by construction. The Continuum Memory System / Hope multi-frequency backbone is taken from Behrouz et al. (2025) (self-citation), but it functions as a reusable substrate on which Sleep is defined, not as the measured target whose correctness is assumed. Table 3 even reports a pure-Transformer Sleep variant, and all gains are measured against independent baselines (ICL, EWC, InCA, SEAL, GRPO, SFT, Cartridges). No step reduces Eq. X to Eq. Y by definition or by an unverified self-citation chain. The architectural entanglement noted by the skeptic is a validity/isolation concern, not circularity under the stated criteria.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 4 invented entities

The central claim rests on Nested Learning’s continuum memory frequencies, the premise that capacity growth plus frozen old parameters blocks interference, and a stack of design choices (expansion rank, sleep schedule, distillation/imitation mix, dream selection) that are free or hand-set. Human sleep stages are motivational analogy, not evidence. Invented process names (Knowledge Seeding, LTI, Sleep) organize the method but are not independently measured entities outside the proposed training loop.

free parameters (6)
  • on-policy mix λ and distillation/LTI mix α
    Control how much student-generated data and how strongly imitation reward vs divergence enter L_KS; chosen as design hyperparameters, not derived.
  • semantic/absolute reward mix γ and Levenshtein threshold z0
    Define the Learning-to-Imitate reward (Eq. 3–4); thresholds and weights are free knobs affecting consolidation quality.
  • low-rank expansion rank d_low and expert count schedule
    How many new parameters unlock per sleep step; capacity and interference trade-off is hand-designed (appendix uses 5 MLP blocks dim 64).
  • memory update frequencies / chunk sizes C^(ℓ) and number of sleep stages
    Define when consolidation fires and how persistent long-term memory is; varied in experiments (Hope-1/2/3) as free architectural schedule.
  • dream count m, top-k importance selection, random sample count b
    Curriculum size and diversity for Dreaming; task-specific (e.g., 5 dreams/passage; ARC 60 sample/45 reject).
  • LoRA rank/alpha and SFT/RL step budgets in Dreaming
    Inner-loop adaptation cost and capacity (Table 5: r=64, α=128); standard PEFT free parameters that affect reported gains.
axioms (6)
  • ad hoc to paper Human NREM/REM sleep functions (consolidation then integration/exploration) are a valid design template for LLM offline phases.
    Used throughout Introduction and §3 as motivation; not empirically required for the algorithms to be well-defined.
  • domain assumption Continuum Memory System: modules with ordered update frequencies form a short-to-long-term memory hierarchy (Nested Learning).
    §2.2 Definition 1 and Eqs. 1–2; substrate for when sleep fires and what is distilled where.
  • domain assumption Catastrophic forgetting is primarily a limited-capacity / interference problem solvable by expanding new experts and freezing old ones.
    §3.2 parameter expansion rationale; load-bearing for claiming consolidation without CF.
  • domain assumption On-policy distillation plus imitation rewards transfer usable knowledge into newly unlocked parameters without needing external labeled sleep data.
    §3.3 KS objective; assumes teacher logits/samples are sufficient supervision for the expanded student.
  • domain assumption Gradient-norm/importance scoring of synthetic dreams plus random-expert injection yields useful self-improvement curricula without destructive drift.
    §3.4 Dreaming; builds on SEAL/ReSTEM but adds selection and novelty assumptions.
  • standard math Standard optimization and PEFT facts (SGD/Adam-style updates, LoRA, MoE routing) behave as usual under the proposed schedule.
    Background ML machinery used without re-derivation.
invented entities (4)
  • Sleep paradigm (wake vs sleep lifecycle for continual learners) no independent evidence
    purpose: Replace train/test framing with periodic offline consolidation and self-improvement.
    Organizing framework of the paper; falsifiable only via the empirical pipeline, not as a physical entity.
  • Knowledge Seeding / Self-Knowledge Seeding (upward distillation into expanded parameters) no independent evidence
    purpose: Transfer high-frequency memory into larger lower-frequency capacity while preserving old weights.
    Core proposed mechanism in §3.3; evidence is internal ablations and task gains only.
  • Learning-to-Imitate (LTI) reward combining semantic and Levenshtein alignment no independent evidence
    purpose: Teach the expanded student not only teacher knowledge but teacher sampling behavior.
    Eqs. 3–4; component of GKD-style objective unique to this design.
  • Dreaming curriculum with gradient-based dream selection and random-expert novelty no independent evidence
    purpose: Self-generate rehearsal data that improves recent knowledge without pure mode collapse.
    §3.4 modifications to SEAL-style self-edits; validated only inside reported experiments.

pith-pipeline@v1.1.0-grok45 · 34378 in / 4299 out tokens · 49216 ms · 2026-07-13T07:43:24.908037+00:00 · methodology

0 comments
read the original abstract

The past few decades have witnessed significant advances in the design of machine learning algorithms, from early studies on task-specific shallow models to more general deep Large Language Models (LLMs). Despite showing promising results in tasks that require instant prediction or in-context learning, existing models lack the ability to continually learn and effectively transfer their temporal in-context knowledge to their long-term parameters. Inspired by human learning process, we introduce a ''Sleep'' paradigm that allows the models to continually learn, distill their short-term fragile memories into stable long-term knowledge with replay, and recursively improve themselves with ''Dreaming'' process. In more detail, sleep consists of two stages: (1) Memory Consolidation: an upward distillation process, called Knowledge Seeding, where the memories of a smaller-self are distilled into a larger network to provide more capacity while preserving the knowledge. As a proof of concept, we present a new Generalized Distillation process for {Knowledge Seeding} (i.e., the combination of on-policy distillation with Reinforcement Learning (RL)-based imitation learning); (2) Dreaming: a self-improvement phase, where the model uses RL to generate a curriculum of synthetic data to rehearse new knowledge and refine existing capabilities without human supervision. Our experiments on long-horizon, continual learning, knowledge incorporation, and few-shot generalization tasks support the importance of the sleep stage.

Figures

Figures reproduced from arXiv: 2606.03979 by Adel Javanmard, Ali Behrouz, Farnoosh Hashemi, Vahab Mirrokni.

Figure 1
Figure 1. Figure 1: (Conventional Machine Learning vs. Continual Learning) While in conventional machine learning often the lifespan of the model is divided to test and training time, continual learning setup does not have these phases. We suggest that a continual learner need to have different stages of activeness in learning, which we refer to as: (i) Active or Wake Time, and (ii) Sleep Time. Sleep time is not a passive sta… view at source ↗
Figure 2
Figure 2. Figure 2: An overview of Memory Consolidation. The model increases its own number of parameters to enhance its [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Class-incremental learning for text classification is evaluated on the (Left) CLINC dataset (Larson et al. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of memory levels on in-context learning performance for (Left) MK-NIAH from RULER (Hsieh et al. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Results on the BABILong benchmark. Red points correspond to fine-tuned models, whereas blue points correspond to zero-shot eval￾uations of large-scale models [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Multi-frequency memory hierarchy. Updates enter the High-Frequency FFN via repeated Parameter Expansion; [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Memory consolidation by routed expert updates. Across Sleep cycles (left [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops

    cs.AI 2026-07 conditional novelty 6.0

    A survey of 1,250 papers organizes AI self-improvement along two axes—what is improved and loop closure—finding that demonstrated self-improvement strength tracks a verification hierarchy from formal verifiers down to...

Reference graph

Works this paper leans on

145 extracted references · 61 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Loss of recent memory after bilateral hippocampal lesions

    William Beecher Scoville and Brenda Milner. “Loss of recent memory after bilateral hippocampal lesions”. In: Journal of neurology, neurosurgery, and psychiatry20.1 (1957), p. 11

    1957

  2. [2]

    1987.url: https://people.idsia.ch/ ~juergen/diploma1987ocr.pdf

    Jürgen Schmidhuber.Evolutionary Principles in Self-Referential Learning. 1987.url: https://people.idsia.ch/ ~juergen/diploma1987ocr.pdf

    1987

  3. [3]

    Self-improving reactive agents based on reinforcement learning, planning and teaching

    Long-Ji Lin. “Self-improving reactive agents based on reinforcement learning, planning and teaching”. In:Machine Learning8.3–4 (1992), pp. 293–321

    1992

  4. [4]

    Learning to control fast-weight memories: An alternative to recurrent nets. Accepted for publication in

    Juergen Schmidhuber. “Learning to control fast-weight memories: An alternative to recurrent nets. Accepted for publication in”. In:Neural Computation(1992)

    1992

  5. [5]

    Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory

    James L. McClelland, Bruce L. McNaughton, and Randall C. O’Reilly. “Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory”. In:Psychological Review102.3 (1995), pp. 419–457

    1995

  6. [6]

    Retrograde amnesia and memory consolidation: a neurobiological perspective

    Larry R Squire and Pablo Alvarez. “Retrograde amnesia and memory consolidation: a neurobiological perspective”. In:Current opinion in neurobiology5.2 (1995), pp. 169–177

    1995

  7. [7]

    Synaptic tagging and long-term potentiation

    Uwe Frey and Richard GM Morris. “Synaptic tagging and long-term potentiation”. In:Nature385.6616 (1997), pp. 533–536

    1997

  8. [8]

    The plastic human brain cortex

    Alvaro Pascual-Leone, Amir Amedi, Felipe Fregni, and Lotfi B Merabet. “The plastic human brain cortex”. In:Annu. Rev. Neurosci.28.1 (2005), pp. 377–401

    2005

  9. [9]

    Sleep-dependent memory consolidation

    Robert Stickgold. “Sleep-dependent memory consolidation”. In:Nature437.7063 (2005), pp. 1272–1278

    2005

  10. [10]

    Reverse replay of behavioural sequences in hippocampal place cells during the awake state

    David J Foster and Matthew A Wilson. “Reverse replay of behavioural sequences in hippocampal place cells during the awake state”. In:Nature440.7084 (2006), pp. 680–683

    2006

  11. [11]

    Sleep function and synaptic homeostasis

    Giulio Tononi and Chiara Cirelli. “Sleep function and synaptic homeostasis”. In:Sleep medicine reviews10.1 (2006), pp. 49–62

    2006

  12. [12]

    Dbpedia: A nucleus for a web of open data

    Sören Auer, Christian Bizer, Georgi Kobilarov, Jens Lehmann, Richard Cyganiak, and Zachary Ives. “Dbpedia: A nucleus for a web of open data”. In:international semantic web conference. Springer. 2007, pp. 722–735

    2007

  13. [13]

    Coordinated memory replay in the visual cortex and hippocampus during sleep

    Daoyun Ji and Matthew A Wilson. “Coordinated memory replay in the visual cortex and hippocampus during sleep”. In:Nature neuroscience10.1 (2007), pp. 100–107

    2007

  14. [14]

    Plasticity in the developing brain: implications for rehabilitation

    Michael V Johnston. “Plasticity in the developing brain: implications for rehabilitation”. In:Developmental disabilities research reviews15.2 (2009), pp. 94–101

    2009

  15. [15]

    Replay of rule-learning related neural patterns in the prefrontal cortex during sleep

    Adrien Peyrache, Mehdi Khamassi, Karim Benchenane, Sidney I Wiener, and Francesco P Battaglia. “Replay of rule-learning related neural patterns in the prefrontal cortex during sleep”. In:Nature neuroscience12.7 (2009), pp. 919–926

    2009

  16. [16]

    Memory, sleep and dreaming: experiencing consolidation

    Erin J Wamsley and Robert Stickgold. “Memory, sleep and dreaming: experiencing consolidation”. In:Sleep medicine clinics6.1 (2011), p. 97

    2011

  17. [17]

    About sleep’s role in memory

    Björn Rasch and Jan Born. “About sleep’s role in memory”. In:Physiological reviews(2013)

    2013

  18. [18]

    The role of sleep in emotional brain function

    Andrea N Goldstein and Matthew P Walker. “The role of sleep in emotional brain function”. In:Annual review of clinical psychology10.1 (2014), pp. 679–708

    2014

  19. [19]

    Distilling the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. “Distilling the knowledge in a neural network”. In:arXiv preprint arXiv:1503.02531(2015)

    Pith/arXiv arXiv 2015

  20. [20]

    Human-level control through deep reinforcement learning

    Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, et al. “Human-level control through deep reinforcement learning”. In:Nature518.7540 (2015), pp. 529–533

    2015

  21. [21]

    Memory consolidation

    Larry R Squire, Lisa Genzel, John T Wixted, and Richard G Morris. “Memory consolidation”. In:Cold Spring Harbor perspectives in biology7.8 (2015), a021766

    2015

  22. [22]

    Sequence-level knowledge distillation

    Yoon Kim and Alexander M Rush. “Sequence-level knowledge distillation”. In:Proceedings of the 2016 conference on empirical methods in natural language processing. 2016, pp. 1317–1327

    2016

  23. [23]

    What learning systems do intelligent agents need? Complementary learning systems theory updated

    Dharshan Kumaran, Demis Hassabis, and James L. McClelland. “What learning systems do intelligent agents need? Complementary learning systems theory updated”. In:Trends in Cognitive Sciences20.7 (2016), pp. 512–534

    2016

  24. [24]

    SQuAD: 100,000+ Questions for Machine Comprehension of Text

    Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. “SQuAD: 100,000+ Questions for Machine Comprehension of Text”. In:Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing. Ed. by Jian Su, Kevin Duh, and Xavier Carreras. Association for Computational Linguistics, 2016.url: https: //aclanthology.org/D16-1264/. 14

    2016

  25. [25]

    Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. “Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks”. In:Proceedings of the 34th International Conference on Machine Learning. Ed. by Doina Precup and Yee Whye Teh. Proceedings of Machine Learning Research. PMLR, 2017.url: https://proceedings.mlr.press/ v70/finn17a.html

    2017

  26. [26]

    Neuroscience-inspired artificial intelligence

    Demis Hassabis, Dharshan Kumaran, Christopher Summerfield, and Matthew Botvinick. “Neuroscience-inspired artificial intelligence”. In:Neuron95.2 (2017), pp. 245–258

    2017

  27. [27]

    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”. In:Proceedings of the national academy of sciences114.13 (2017), pp. 3521–3526

    2017

  28. [28]

    REM sleep selectively prunes and maintains new synapses in development and learning

    Wei Li, Lei Ma, Guang Yang, and Wen-Biao Gan. “REM sleep selectively prunes and maintains new synapses in development and learning”. In:Nature neuroscience20.3 (2017), pp. 427–437

    2017

  29. [29]

    Attention is All you Need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. “Attention is All you Need”. In:Advances in Neural Information Processing Systems. Ed. by I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett. Vol. 30. Cur- ran Associates, Inc., 2017.url: htt...

    2017

  30. [30]

    Recurrent world models facilitate policy evolution

    David Ha and Jürgen Schmidhuber. “Recurrent world models facilitate policy evolution”. In:Advances in Neural Information Processing Systems (NeurIPS). Vol. 31. 2018, pp. 2451–2463

    2018

  31. [31]

    World models

    David Ha and Jürgen Schmidhuber. “World models”. In:arXiv preprint arXiv:1803.101222.3 (2018), p. 440

    Pith/arXiv arXiv 2018

  32. [32]

    Measuring catastrophic forgetting in neural networks

    Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. “Measuring catastrophic forgetting in neural networks”. In:Proceedings of the AAAI conference on artificial intelligence. Vol. 32. 1. 2018

    2018

  33. [33]

    An Evaluation Dataset for Intent Classification and Out-of-Scope Prediction

    Stefan Larson, Anish Mahendran, Joseph J Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K Kummerfeld, Kevin Leach, Michael A Laurenzano, Lingjia Tang, et al. “An Evaluation Dataset for Intent Classification and Out-of-Scope Prediction”. In:Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th Inte...

    2019

  34. [34]

    Fast transformer decoding: One write-head is all you need

    Noam Shazeer. “Fast transformer decoding: One write-head is all you need”. In:arXiv preprint arXiv:1911.02150 (2019)

    Pith/arXiv arXiv 1911

  35. [35]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. “Language models are few-shot learners”. In:Advances in neural information processing systems33 (2020), pp. 1877–1901

    2020

  36. [36]

    Efficient Intent Detection with Dual Sentence Encoders

    Inigo Casanueva, Tadas Temcinas, Daniela Gerz, Matthew Henderson, and Ivan Vulic. “Efficient Intent Detection with Dual Sentence Encoders”. In:ACL 2020(2020), p. 38

    2020

  37. [37]

    Can sleep protect memories from catastrophic forgetting?

    Oscar C. González, Yury Sokolov, Giri P. Krishnan, Jean Erik Delanois, and Maxim Bazhenov. “Can sleep protect memories from catastrophic forgetting?” In:eLife9 (2020), e51005

    2020

  38. [38]

    Dream to control: Learning behaviors by latent imagination

    Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. “Dream to control: Learning behaviors by latent imagination”. In:International Conference on Learning Representations (ICLR). 2020

    2020

  39. [39]

    Transformers are rnns: Fast au- toregressive transformers with linear attention

    Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. “Transformers are rnns: Fast au- toregressive transformers with linear attention”. In:International conference on machine learning. PMLR. 2020, pp. 5156–5165

    2020

  40. [40]

    A dataset of information- seeking questions and answers anchored in research papers

    Pradeep Dasigi, Kyle Lo, Iz Beltagy, Arman Cohan, Noah A Smith, and Matt Gardner. “A dataset of information- seeking questions and answers anchored in research papers”. In:arXiv preprint arXiv:2105.03011(2021)

    Pith/arXiv arXiv 2021

  41. [41]

    Stepwise synaptic plasticity events drive the early phase of memory consolidation

    Akihiro Goto, Ayaka Bota, Ken Miya, Jingbo Wang, Suzune Tsukamoto, Xinzhi Jiang, Daichi Hirai, Masanori Murayama, Tomoki Matsuda, Thomas J. McHugh, Takeharu Nagai, and Yasunori Hayashi. “Stepwise synaptic plasticity events drive the early phase of memory consolidation”. In:Science374.6569 (2021), pp. 857–863.doi: 10.1126/science.abj9195 . eprint: https://...

    doi:10.1126/science.abj9195 2021

  42. [42]

    McGraw-Hill, 2021

    Eric R Kandell, Jojhn D Koester, Sarah H Mack, and Steven Siegelbaum.Principles of neural science. McGraw-Hill, 2021

    2021

  43. [43]

    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: arXiv preprint arXiv:2104.08691(2021)

    Pith/arXiv arXiv 2021

  44. [44]

    Prefix-Tuning: Optimizing Continuous Prompts for Generation

    Xiang Lisa Li and Percy Liang. “Prefix-Tuning: Optimizing Continuous Prompts for Generation”. In:Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). Ed. by Chengqing Zong, Fei Xia, Wenjie Li, and Roberto 15 Navigli. Onlin...

    doi:10.18653/v1/2021.acl- 2021

  45. [45]

    What learning algorithm is in-context learning? investigations with linear models

    Ekin Akyürek, Dale Schuurmans, Jacob Andreas, Tengyu Ma, and Denny Zhou. “What learning algorithm is in-context learning? investigations with linear models”. In:arXiv preprint arXiv:2211.15661(2022)

    Pith/arXiv arXiv 2022

  46. [46]

    Recurrent memory transformer

    Aydar Bulatov, Yury Kuratov, and Mikhail Burtsev. “Recurrent memory transformer”. In:Advances in Neural Information Processing Systems35 (2022), pp. 11079–11091

    2022

  47. [47]

    LoRA: Low-Rank Adaptation of Large Language Models

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. “LoRA: Low-Rank Adaptation of Large Language Models”. In:International Conference on Learning Representations. 2022.url:https://openreview.net/forum?id=nZeVKeeFYf9

    2022

  48. [48]

    A modern self-referential weight matrix that learns to modify itself

    Kazuki Irie, Imanol Schlag, Róbert Csordás, and Jürgen Schmidhuber. “A modern self-referential weight matrix that learns to modify itself”. In:International Conference on Machine Learning. PMLR. 2022.url: https://proceedings. mlr.press/v162/irie22b.html

    2022

  49. [49]

    Sleep-like unsupervised replay reduces catastrophic forgetting in artificial neural networks

    Timothy Tadros, Giri P. Krishnan, Ramyaa Ramyaa, and Maxim Bazhenov. “Sleep-like unsupervised replay reduces catastrophic forgetting in artificial neural networks”. In:Nature Communications13.1 (2022), p. 7742

    2022

  50. [50]

    STaR: Bootstrapping Reasoning With Reasoning

    Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. “STaR: Bootstrapping Reasoning With Reasoning”. In: Advances in Neural Information Processing Systems. Ed. by S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh. Curran Associates, Inc., 2022.url: https://proceedings.neurips.cc/paper_files/paper/2022/ file/639a9a172c044fbb64175b5fad42e9a...

    2022

  51. [51]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. “Gpt-4 technical report”. In:arXiv preprint arXiv:2303.08774(2023)

    Pith/arXiv arXiv 2023

  52. [52]

    Gqa: Training generalized multi-query transformer models from multi-head checkpoints

    Joshua Ainslie, James Lee-Thorp, Michiel De Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. “Gqa: Training generalized multi-query transformer models from multi-head checkpoints”. In:arXiv preprint arXiv:2305.13245(2023)

    Pith/arXiv arXiv 2023

  53. [53]

    Adapting language models to compress contexts

    Alexis Chevalier, Alexander Wettig, Anirudh Ajith, and Danqi Chen. “Adapting language models to compress contexts”. In:arXiv preprint arXiv:2305.14788(2023)

    Pith/arXiv arXiv 2023

  54. [54]

    In-context autoencoder for context compression in a large language model

    Tao Ge, Jing Hu, Lei Wang, Xun Wang, Si-Qing Chen, and Furu Wei. “In-context autoencoder for context compression in a large language model”. In:arXiv preprint arXiv:2307.06945(2023)

    Pith/arXiv arXiv 2023

  55. [55]

    Mixture of cluster-conditional lora experts for vision-language instruction tuning

    Yunhao Gou, Zhili Liu, Kai Chen, Lanqing Hong, Hang Xu, Aoxue Li, Dit-Yan Yeung, James T Kwok, and Yu Zhang. “Mixture of cluster-conditional lora experts for vision-language instruction tuning”. In:arXiv preprint arXiv:2312.12379(2023)

    Pith/arXiv arXiv 2023

  56. [56]

    Mamba: Linear-time sequence modeling with selective state spaces

    Albert Gu and Tri Dao. “Mamba: Linear-time sequence modeling with selective state spaces”. In:arXiv preprint arXiv:2312.00752(2023)

    Pith/arXiv arXiv 2023

  57. [57]

    Lorahub: Efficient cross-task generalization via dynamic lora composition

    Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du, and Min Lin. “Lorahub: Efficient cross-task generalization via dynamic lora composition”. In:arXiv preprint arXiv:2307.13269(2023)

    Pith/arXiv arXiv 2023

  58. [58]

    Llmlingua: Compressing prompts for accelerated inference of large language models

    Huiqiang Jiang, Qianhui Wu, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. “Llmlingua: Compressing prompts for accelerated inference of large language models”. In:arXiv preprint arXiv:2310.05736(2023)

    Pith/arXiv arXiv 2023

  59. [59]

    Unlocking context constraints of llms: Enhancing context efficiency of llms with self-information-based content filtering

    Yucheng Li. “Unlocking context constraints of llms: Enhancing context efficiency of llms with self-information-based content filtering”. In:arXiv preprint arXiv:2304.12102(2023)

    Pith/arXiv arXiv 2023

  60. [60]

    CodeGen: An Open Large Language Model for Code with Multi-Turn Program Synthesis

    Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, and Caiming Xiong. “CodeGen: An Open Large Language Model for Code with Multi-Turn Program Synthesis”. In:The Eleventh Interna- tional Conference on Learning Representations. 2023.url:https://openreview.net/forum?id=iaYcJKpY2B_

    2023

  61. [61]

    Memgpt: Towards llms as operating systems

    Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G Patil, Ion Stoica, and Joseph E Gonzalez. “Memgpt: Towards llms as operating systems”. In:arXiv preprint arXiv:2310.08560(2023)

    Pith/arXiv arXiv 2023

  62. [62]

    Are emergent abilities of large language models a mirage?

    Rylan Schaeffer, Brando Miranda, and Sanmi Koyejo. “Are emergent abilities of large language models a mirage?” In:Advances in neural information processing systems36 (2023), pp. 55565–55581

    2023

  63. [63]

    Visionllm: Large language model is also an open-ended decoder for vision-centric tasks

    Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. “Visionllm: Large language model is also an open-ended decoder for vision-centric tasks”. In:Advances in Neural Information Processing Systems36 (2023), pp. 61501–61513

    2023

  64. [64]

    H2o: Heavy-hitter oracle for efficient generative inference of large language models

    Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. “H2o: Heavy-hitter oracle for efficient generative inference of large language models”. In:Advances in Neural Information Processing Systems36 (2023), pp. 34661–34710. 16

    2023

  65. [65]

    Phi-4 technical report

    Marah Abdin, Jyoti Aneja, Harkirat Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Rus- sell J Hewett, Mojan Javaheripi, Piero Kauffmann, et al. “Phi-4 technical report”. In:arXiv preprint arXiv:2412.08905 (2024)

    Pith/arXiv arXiv 2024

  66. [66]

    On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes

    Rishabh Agarwal, Nino Vieillard, Yongchao Zhou, Piotr Stanczyk, Sabela Ramos Garea, Matthieu Geist, and Olivier Bachem. “On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes”. In:The Twelfth International Conference on Learning Representations. 2024.url: https://openreview.net/forum?id= 3zKtaqxLhW

    2024

  67. [67]

    The Surprising Effectiveness of Test-Time Training for Few-Shot Learning

    Ekin Akyürek, Mehul Damani, Adam Zweiger, Linlu Qiu, Han Guo, Jyothish Pari, Yoon Kim, and Jacob Andreas. “The Surprising Effectiveness of Test-Time Training for Few-Shot Learning”. In:Forty-second International Conference on Machine Learning. 2024

    2024

  68. [68]

    In-context language learning: Architectures and algorithms

    Ekin Akyürek, Bailin Wang, Yoon Kim, and Jacob Andreas. “In-context language learning: Architectures and algorithms”. In:arXiv preprint arXiv:2401.12973(2024)

    Pith/arXiv arXiv 2024

  69. [69]

    Simple linear attention language models balance the recall-throughput tradeoff

    Simran Arora, Sabri Eyuboglu, Michael Zhang, Aman Timalsina, Silas Alberti, James Zou, Atri Rudra, and Christo- pher Re. “Simple linear attention language models balance the recall-throughput tradeoff”. In:Forty-first International Conference on Machine Learning. 2024.url:https://openreview.net/forum?id=e93ffDcpH3

    2024

  70. [70]

    xLSTM: Extended Long Short-Term Memory

    Maximilian Beck, Korbinian Pöppel, Markus Spanring, Andreas Auer, Oleksandra Prudnikova, Michael Kopp, Günter Klambauer, Johannes Brandstetter, and Sepp Hochreiter. “xLSTM: Extended Long Short-Term Memory”. In: arXiv preprint arXiv:2405.04517(2024)

    Pith/arXiv arXiv 2024

  71. [71]

    Titans: Learning to memorize at test time

    Ali Behrouz, Peilin Zhong, and Vahab Mirrokni. “Titans: Learning to memorize at test time”. In:arXiv preprint arXiv:2501.00663(2024)

    Pith/arXiv arXiv 2024

  72. [72]

    Dated Data: Tracing Knowledge Cutoffs in Large Language Models

    Jeffrey Cheng, Marc Marone, Orion Weller, Dawn Lawrie, Daniel Khashabi, and Benjamin Van Durme. “Dated Data: Tracing Knowledge Cutoffs in Large Language Models”. In:First Conference on Language Modeling. 2024.url: https://openreview.net/forum?id=wS7PxDjy6m

    2024

  73. [73]

    A Survey on In-context Learning

    Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Baobao Chang, Xu Sun, Lei Li, and Zhifang Sui. “A Survey on In-context Learning”. In:Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. Ed. by Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen. Miami, Florida, USA: Asso...

    doi:10.18653/v1/ 2024

  74. [74]

    The llama 3 herd of models

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. “The llama 3 herd of models”. In:arXiv e-prints(2024), arXiv–2407

    2024

  75. [75]

    RULER: What’s the Real Context Size of Your Long-Context Language Models?

    Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, and Boris Ginsburg. “RULER: What’s the Real Context Size of Your Long-Context Language Models?” In:First Conference on Language Modeling. 2024.url:https://openreview.net/forum?id=kIoBbc76Sy

    2024

  76. [76]

    Simple and Scalable Strategies to Continually Pre-train Large Language Models

    Adam Ibrahim, Benjamin Thérien, Kshitij Gupta, Mats Leon Richter, Quentin Gregory Anthony, Eugene Belilovsky, Timothée Lesort, and Irina Rish. “Simple and Scalable Strategies to Continually Pre-train Large Language Models”. In:Transactions on Machine Learning Research(2024).issn: 2835-8856.url: https://openreview.net/forum?id= DimPeeCxKO

    2024

  77. [77]

    Knowledge injection via prompt distillation

    Kalle Kujanpää, Harri Valpola, and Alexander Ilin. “Knowledge injection via prompt distillation”. In:arXiv preprint arXiv:2412.14964(2024)

    Pith/arXiv arXiv 2024

  78. [78]

    Babilong: Testing the limits of llms with long context reasoning-in-a-haystack

    Yury Kuratov, Aydar Bulatov, Petr Anokhin, Ivan Rodkin, Dmitry Sorokin, Artyom Sorokin, and Mikhail Burtsev. “Babilong: Testing the limits of llms with long context reasoning-in-a-haystack”. In:Advances in Neural Information Processing Systems37 (2024), pp. 106519–106554

    2024

  79. [79]

    Mixlora: Enhancing large language models fine-tuning with lora-based mixture of experts

    Dengchun Li, Yingzi Ma, Naizheng Wang, Zhengmao Ye, Zhiyuan Cheng, Yinghao Tang, Yan Zhang, Lei Duan, Jie Zuo, Cal Yang, et al. “Mixlora: Enhancing large language models fine-tuning with lora-based mixture of experts”. In:arXiv preprint arXiv:2404.15159(2024)

    Pith/arXiv arXiv 2024

  80. [80]

    Snapkv: Llm knows what you are looking for before generation

    Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. “Snapkv: Llm knows what you are looking for before generation”. In:Advances in Neural Information Processing Systems37 (2024), pp. 22947–22970

    2024

Showing first 80 references.