Pith. sign in

REVIEW 4 major objections 5 minor 45 references

UniMem claims that a frozen LLM can expand its own memory on the fly by routing each query to either a retrieval buffer or a trainable parametric unit, and that this autonomous coordination lifts exact-match accuracy by 4.0 points on averag

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 · deepseek-v4-flash

2026-08-01 00:49 UTC pith:S4WKF53T

load-bearing objection Credible system paper whose headline gain depends on unmeasured cluster quality; deserves refereeing but needs code and cluster-level analysis. the 4 major comments →

arxiv 2607.26017 v1 pith:S4WKF53T submitted 2026-07-28 cs.CL

UniMem: Complementary Episodic-to-Parametric Memory for Boundary-Agnostic Task Streams

Siyu Xia , Chenheng Zhang , Yanting Wu , Haoxuan Li , Jiajun Chai , Xiaohan Wang , Guojun Yin , Wei Lin
show 3 more authors
Zhouchen Lin Haifeng Zhang Jun Wang
This is my paper
classification cs.CL
keywords episodic-to-parametric memoryboundary-agnostic task streamsrouting tokensmemory consolidationstability-plasticity dilemmaComplementary Learning SystemsNormalized Compression DistanceProcedural KV Memory
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.

The paper tries to establish that a single frozen language model can handle boundary-agnostic, evolving task streams by autonomously managing two complementary memory pathways: an episodic buffer for novel or sparse tasks and expandable parametric memory blocks for recurring patterns. The key idea is to use learnable routing tokens as memory controllers that decide, per query, whether to retrieve evidence from a buffer or to activate a task-specific Procedural KV memory. If correct, this would let agents acquire new tasks during deployment without task labels, without forgetting, and without uncontrolled parameter growth. The reported experimental gain of 4.0 exact-match points across three backbones is the concrete payoff the paper uses to support this claim.

Core claim

On its own terms, UniMem demonstrates that the stability-plasticity dilemma in streaming LLM deployment can be managed by physically separating task identification from task execution. Each incoming query is scored against learned routing tokens plus a novelty sentinel; only high-confidence matches activate a parametric memory unit, while anything else is sent to an episodic buffer served by retrieval. When the buffer fills, an unsupervised pipeline—using Normalized Compression Distance and HDBSCAN clustering—groups recurring samples into candidate tasks, filters them by outlier removal and cluster cohesion, and promotes only qualifying clusters into new routing tokens paired with Procedural

What carries the argument

The central mechanism is the routing token matrix E = [e_1, ..., e_K, e_NOVEL], where each known-task token e_k is paired with a pluggable parametric memory block P_k, instantiated here as layer-wise learnable key-value slots injected into the frozen transformer's cross-attention with learned gates. The novelty sentinel e_NOVEL is calibrated during initialization to compete with known tokens, and a confidence rule (p* > p0 and p* ≥ τ, with τ = 0.7) decides whether to activate a parametric unit or send the query to the episodic buffer. Task discovery relies on NCD as a string-compression similarity measure, HDBSCAN for label-free clustering, and a strict quality gate before consolidation. Thi

Load-bearing premise

The entire discovery pipeline assumes that a recurring task leaves stable, surface-level textual traces—stable instruction and query phrasing—that show up as compressible patterns in gzip-based NCD; if the same procedure is phrased differently each time or evolves semantically without lexical overlap, clustering will fail or misconsolidate, breaking the central improvement.

What would settle it

Build a streaming sequence where each task's instruction is rewritten every time it appears using different synonyms and sentence structures, so that samples from the same latent task share semantics but not surface text. If UniMem's routing accuracy drops, few or no clusters pass the NCD cohesion threshold, and exact-match scores fall back toward the retrieval-only baseline, then the claim that recurring patterns can be discovered without labels from text would be disproven.

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

If this is right

  • A deployed LLM agent can accumulate task-specific skills without any task-boundary supervision, as long as recurring tasks leave surface-level textual traces.
  • Parameter growth becomes proportional to the number of genuinely recurring task patterns, not to the number of tasks or samples, because sparse and one-off tasks stay in the episodic buffer.
  • The routing-token mechanism remains reliable at scale: routing accuracy stays above 85% for LLaMA-3.1-8B even with 100 tasks, so the decoupled architecture does not degrade as memory expands.
  • Task-specific memory blocks reduce interference compared to a shared memory block, and the advantage grows with task count (EM gap widens from +1.2% at 10 tasks to +2.7% at 50 tasks).
  • The framework is modular in its execution memory: any PEFT-style module could replace the Procedural KV Memory, so the routing-plus-consolidation lifecycle is a general recipe for autonomous memory expansion.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural stress test the paper leaves implicit: if the same task is repeatedly rephrased so that instruction and query strings share little lexical overlap while semantics stay aligned, NCD-based clustering should fail to discover the recurrence, and consolidation would either stall or merge distinct tasks. Running such an adversarial stream would directly expose the dependence on surface-form si
  • The reported 4.0-point exact-match gain averages over three backbones; the gain is uneven (e.g., on Qwen3-8B at 100 tasks the improvement over TOKMEM is about 2.7 points, while on LLaMA-3.2-3B it is about 7.0 points). A practitioner choosing a backbone should expect the benefit to depend on the base model's instruction-following strength.
  • The conservative consolidation policy (minimum cluster size 50, strict cutoff τ_cluster = 0.85) implies a latency trade-off: sparse tasks accept degraded retrieval-based performance for many rounds before enough evidence accumulates. Adaptive thresholds based on stream statistics could accelerate consolidation for medium-frequency tasks without a wholesale rewrite of the mechanism.
  • Because the routing tokens and memory blocks are trained separately, one could reuse a routing layer across multiple backbones or transfer consolidated memory units to a new frozen backbone, potentially enabling portable task expertise across model versions.

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

4 major / 5 minor

Summary. The paper proposes UniMem, a framework for LLM memory management in boundary-agnostic streaming task streams. UniMem uses learnable routing tokens to dispatch queries either to task-specific parametric ``Procedural KV'' memory blocks or to an episodic buffer served by retrieval. When the buffer fills, NCD and HDBSCAN are used to cluster buffered queries without task labels; clusters passing a quality gate are consolidated into new routing tokens and parametric memory units via SFT. The paper reports experiments on SNI streaming sequences and a SuperGLUE mixed stream, claiming an average gain of 4.0 EM points across three backbone models and demonstrating autonomous creation of 76 parametric units over 100 novel tasks. The manuscript includes detailed appendices on routing, clustering, baselines, computational budget, and limitations.

Significance. If the results hold, UniMem addresses a practically relevant problem: enabling a frozen LLM to autonomously expand parametric memory in a task-label-free streaming setting while retaining episodic coverage for sparse tasks. The architecture is clearly motivated by complementary learning systems, and the decoupling of routing from execution is a sensible design that goes beyond single-token procedural memory. The paper is strong on protocol transparency: streaming construction, baseline implementations, hyperparameters, and computational budgets are documented in detail, and the arithmetic behind the abstract's 4.0 EM claim is consistent with Table 1. The work would be a useful contribution to continual learning and agent memory, but the empirical evidence as presented is not yet sufficient to rule out that the central consolidation mechanism relies on favorable clustering outcomes that are never directly measured.

major comments (4)
  1. [§4.2, Fig. 5; App. E.6, Eq. (22)] The central claim that recurring tasks are correctly discovered and consolidated is not directly evaluated. Fig. 5 reports that 76 new units are created for 100 novel tasks, but no cluster purity, precision/recall, confusion matrix, or per-unit error is given. If clusters are impure, SFT will bake mixed task procedures into a single memory block; if clusters fragment, the method silently degenerates to episodic RAG. Both failure modes could leave the aggregate EM numbers unaffected while undermining the intended mechanism. Please report clustering quality against ground-truth task labels and, at minimum, per-consolidation EM changes. Without this, the headline gain is not interpretable as evidence of successful consolidation.
  2. [App. A, Eq. (8); App. E.4–E.6] Several hand-set thresholds control the entire episodic-to-parametric lifecycle: τroute = 0.7, τcluster = 0.85, k = 0.5, Smin = 50, min_samples = 100, and cache capacity C = 1600. No sensitivity analysis is provided for any of them. A method whose behavior hinges on these values needs at least a local robustness study (e.g., varying τcluster over [0.7, 0.95] and τroute over [0.5, 0.9]) to show that the 4.0 EM gain is not an artifact of a finely tuned operating point.
  3. [Table 1 and §4.1] Statistical support for the headline comparison is weak. Table 1 is averaged over three runs using the same seed-42 data split, but no standard deviations, confidence intervals, or significance tests are reported. The advantages over the strongest baseline are only +2.44 EM (LLaMA-3.1-8B) and +3.37 EM (Qwen3-8B); with n=3 same-split runs these could easily be within run-to-run noise. Please report variance across seeds and, where feasible, multiple stream orderings.
  4. [App. E.2–E.3; §1] The task-discovery step relies on surface-form text: xi truncates the instruction and query and NCD is computed at character level. On SNI, tasks share exact instruction strings, so NCD can separate tasks by lexical identity; this partly mitigates the concern that no semantic generalization is needed. However, the paper claims deployment over evolving task streams. If instructions are paraphrased or tasks evolve lexically while preserving procedure, NCD-based clustering may fragment or misgroup. Please test this explicitly, e.g., with rephrased instructions or a synthetic procedure-evolution stream. This is a correctness-risk concern, not a circularity claim: the execution mechanism may be sound even if the discovery step is evaluated only in the easy lexical-identity regime.
minor comments (5)
  1. [App. A and §3.2] Notation for the number of routing tokens varies: T in Eq. (5) versus K in Eq. (1) and Section 3.2. Please unify.
  2. [Table 3] The "w/o RAG cache" rows use em-dashes for the 10-task setting; the caption explains this, but it would be clearer to write "n/a" and note why in the table note.
  3. [Fig. 4] The bars for per-task KV and shared KV are labeled only in the legend; adding values directly on the bars would make the +1.2/+1.9 and +2.7/+2.9 points easier to verify.
  4. [References] Several preprint references are incomplete or inconsistent (e.g., Hu et al. lists partial author groups; "TOKMEM" is rendered as "TOK-MEM" in §2.2). Please normalize.
  5. [Availability] No code or data release is mentioned. Given the complexity of the pipeline, releasing code and configuration files would substantially aid reproducibility.

Circularity Check

0 steps flagged

No significant circularity: central claims are empirical benchmark measurements, not fitted predictions.

full rationale

UniMem's central claims are direct empirical comparisons on SNI and SuperGLUE streams (Tables 1–2, Figure 5), not quantities derived from or fitted to the target result. The routing probability in Eq. 1 is trained with cross-entropy; the Procedural KV Memory in Eq. 2 is trained with a task loss; the NCD/HDBSCAN clustering in Eqs. 14–22 produces clusters that are then SFT-trained. No equation is constructed so that its output equals its input. The observation that sparse tasks remain episodic is a stated, designed consequence of setting Nsparse=10 below HDBSCAN's min_cluster_size=50, and the paper explicitly says this 'tests whether UniMem can avoid premature parameter allocation,' so it is not presented as a surprise or a fitted prediction. The only self-citation is Zhang et al. 2026 in the introduction, which is not load-bearing; the procedural-token foundation is attributed to Wu et al. 2026 (TOKMEM), which is not authored by the present authors. There is no uniqueness-theorem argument, no renamed fitted parameter, and no ansatz smuggled in via self-citation. Unmeasured cluster purity is a correctness risk, but it is not circularity. The paper is self-contained against external benchmarks and no specific reduction or fit-to-target can be quoted. Therefore no significant circularity is found.

Axiom & Free-Parameter Ledger

8 free parameters · 4 axioms · 3 invented entities

The central claim rests on several hand-set thresholds and surface-text clustering; however, none of these are fitted in a way that defines the target result, so circularity burden is low. The architecture adds three trainable components whose utility is evidenced only by in-paper comparisons.

free parameters (8)
  • routing threshold τroute = 0.7
    Appendix A: known memory activated only if top known-token probability exceeds both the sentinel and 0.7; no sensitivity analysis.
  • cluster quality threshold τcluster = 0.85
    Appendix E.6: NCD cohesion threshold for accepting a cluster; fixed without sensitivity analysis.
  • HDBSCAN min_cluster_size Smin = 50
    Appendix E.4: sparse tasks with 10 samples cannot consolidate; drives the episodic-vs-parametric split.
  • HDBSCAN min_samples = 100
    Appendix E.4: conservative cluster assignment; no sensitivity analysis.
  • outlier filter k = 0.5
    Appendix E.5 Eq. (20): threshold for removing samples from candidate clusters.
  • episodic buffer capacity C = 1600
    Appendix E.1: triggers clustering round when cache reaches 1600 samples.
  • initial known/calibration tasks Nknown, Ncalib = 20 / 20
    Appendix C.2: Phase-I routing initialization and novelty calibration sizes.
  • sparse-task ratio rsparse / Nsparse = 10% / 10 samples
    Appendix C.4: long-tail simulation; sparse tasks placed below HDBSCAN cluster minimum.
axioms (4)
  • domain assumption The streaming task distribution is composed of discrete, recurring latent task patterns that can be discovered by clustering.
    Phase II (§3.3) assumes buffered queries form clusters corresponding to tasks; HDBSCAN requires recurring density. If tasks evolve gradually, clusters will not form.
  • domain assumption Task identity is recoverable from surface text: xi = Trunc(I_i)||Trunc(q_i) with gzip NCD captures procedural similarity.
    Appendix E.2–E.3 constructs clustering features purely from instruction and query character strings; no learned semantic representation is used in NCD. Same-procedure-different-wording tasks would be missed.
  • domain assumption A frozen backbone can be adapted by injecting per-task key-value memory via cross-attention without corrupting routing.
    Eq. (2) and Appendix B assume the Procedural KV Memory adds useful task-specific signal while routing tokens stay discriminative; ablation w/o KV gate shows it is critical, but there is no theoretical guarantee.
  • ad hoc to paper Initializing eNOVEL near the centroid of known routing tokens (Eq. 4) creates a well-calibrated novelty signal.
    The Gaussian perturbation and renormalization are a design choice without a principled justification; routing quality depends on it.
invented entities (3)
  • Routing tokens e_k no independent evidence
    purpose: Learnable task selectors that determine whether a query uses an existing parametric memory block or the novelty pathway.
    Architectural component; only in-paper experiments support utility; no external falsifiable handle.
  • Novelty sentinel eNOVEL no independent evidence
    purpose: Competes with known tokens to route novel/sparse queries to the episodic buffer.
    Initialized near token centroid (Eq. 4); calibration uses held-out tasks; no independent evidence outside the paper.
  • Procedural KV memory block P_k no independent evidence
    purpose: Per-task layer-wise key-value slots injected into frozen attention to encode task execution.
    Implemented as cross-attention memory; support comes only from in-paper ablations (Fig. 4, w/o KV gate).

pith-pipeline@v1.3.0-alltime-deepseek · 18040 in / 14156 out tokens · 121737 ms · 2026-08-01T00:49:59.073299+00:00 · methodology

0 comments
read the original abstract

Memory is essential for LLM agents to accumulate task experience and reuse task-specific execution strategies. However, real-world deployment over boundary-agnostic and evolving task streams exposes a fundamental stability-plasticity dilemma. External retrieval-based memory can rapidly absorb new evidence, but it often fails to internalize recurring execution patterns and incurs inference-time retrieval overhead. Parametric memory enables stable and efficient execution once learned, but typically relies on explicit task boundaries and fixed parameter budgets. Inspired by the human brain, which balances plasticity and stability through complementary episodic storage and gradual consolidation, we propose UniMem, a self-routing framework for autonomous memory management. UniMem uses learnable routing tokens as memory controllers, enabling adaptive coordination between complementary memory pathways: novel or sparse tasks are retained in an episodic buffer for retrieval-augmented execution, while recurring and reliable patterns are consolidated into expandable parametric memory. By decoupling task identification from task execution with routing tokens and parametric memory blocks, UniMem expands memory on demand without task labels during deployment or uncontrolled parameter growth. Experiments on long-horizon streaming task sequences show that UniMem consistently outperforms baselines while maintaining execution fidelity, achieving an average gain of 4.0 EM points across three backbone models.

Figures

Figures reproduced from arXiv: 2607.26017 by Chenheng Zhang, Guojun Yin, Haifeng Zhang, Haoxuan Li, Jiajun Chai, Jun Wang, Siyu Xia, Wei Lin, Xiaohan Wang, Yanting Wu, Zhouchen Lin.

Figure 1
Figure 1. Figure 1: The episodic-to-parametric conceptual workflow of [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of UniMem. The left panel shows the overall self-routing architecture for assigning incoming [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Routing token accuracy on SNI streaming task [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study of per-task vs. shared KV [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

45 extracted references

  1. [1]

    2026 , eprint=

    TokMem: One-Token Procedural Memory for Large Language Models , author=. 2026 , eprint=

    2026

  2. [2]

    2023 , eprint=

    Reflexion: Language Agents with Verbal Reinforcement Learning , author=. 2023 , eprint=

    2023

  3. [3]

    2024 , eprint=

    ExpeL: LLM Agents Are Experiential Learners , author=. 2024 , eprint=

    2024

  4. [4]

    2025 , eprint=

    A-MEM: Agentic Memory for LLM Agents , author=. 2025 , eprint=

    2025

  5. [5]

    2025 , eprint=

    MEM1: Learning to Synergize Memory and Reasoning for Efficient Long-Horizon Agents , author=. 2025 , eprint=

    2025

  6. [6]

    2025 , eprint=

    MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-based Memory Agent , author=. 2025 , eprint=

    2025

  7. [7]

    2025 , eprint=

    Zep: A Temporal Knowledge Graph Architecture for Agent Memory , author=. 2025 , eprint=

    2025

  8. [8]

    2025 , eprint=

    HopRAG: Multi-Hop Reasoning for Logic-Aware Retrieval-Augmented Generation , author=. 2025 , eprint=

    2025

  9. [9]

    2025 , eprint=

    G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems , author=. 2025 , eprint=

    2025

  10. [10]

    2025 , eprint=

    From Experience to Strategy: Empowering LLM Agents with Trainable Graph Memory , author=. 2025 , eprint=

    2025

  11. [11]

    Transformer Feed-Forward Layers Are Key-Value Memories

    Geva, Mor and Schuster, Roei and Berant, Jonathan and Levy, Omer. Transformer Feed-Forward Layers Are Key-Value Memories. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021. doi:10.18653/v1/2021.emnlp-main.446

    doi:10.18653/v1/2021.emnlp-main.446 2021

  12. [12]

    Guo, Daya and Yang, Dejian and Zhang, Haowei and Song, Junxiao and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Zhang, Ruoyu and Ma, Shirong and Bi, Xiao and Zhang, Xiaokang and Yu, Xingkai and Wu, Yu and Wu, Z. F. and Gou, Zhibin and Shao, Zhihong and Li, Zhuoshu and Gao, Ziyi and Liu, Aixin and Xue, Bing and Wang, Bingxuan and Wu, Bochao and Feng, Bei ...

  13. [13]

    2023 , eprint=

    GPT Understands, Too , author=. 2023 , eprint=

    2023

  14. [14]

    2022 , eprint=

    P-Tuning v2: Prompt Tuning Can Be Comparable to Fine-tuning Universally Across Scales and Tasks , author=. 2022 , eprint=

    2022

  15. [15]

    2021 , eprint=

    The Power of Scale for Parameter-Efficient Prompt Tuning , author=. 2021 , eprint=

    2021

  16. [16]

    2021 , eprint=

    Prefix-Tuning: Optimizing Continuous Prompts for Generation , author=. 2021 , eprint=

    2021

  17. [17]

    2021 , eprint=

    LoRA: Low-Rank Adaptation of Large Language Models , author=. 2021 , eprint=

    2021

  18. [18]

    2022 , eprint=

    IDPG: An Instance-Dependent Prompt Generation Method , author=. 2022 , eprint=

    2022

  19. [19]

    2024 , eprint=

    IAPT: Instruction-Aware Prompt Tuning for Large Language Models , author=. 2024 , eprint=

    2024

  20. [20]

    2025 , eprint=

    All You Need is One: Capsule Prompt Tuning with a Single Vector , author=. 2025 , eprint=

    2025

  21. [21]

    2022 , eprint=

    Learning to Prompt for Continual Learning , author=. 2022 , eprint=

    2022

  22. [22]

    2024 , eprint=

    MEMORYLLM: Towards Self-Updatable Large Language Models , author=. 2024 , eprint=

    2024

  23. [23]

    2023 , eprint=

    Progressive Prompts: Continual Learning for Language Models , author=. 2023 , eprint=

    2023

  24. [24]

    Mixture-of-LoRAs: An Efficient Multitask Tuning Method for Large Language Models , booktitle =

    Wenfeng Feng and Chuzhan Hao and Yuewei Zhang and Yu Han and Hao Wang , year =. Mixture-of-LoRAs: An Efficient Multitask Tuning Method for Large Language Models , booktitle =

  25. [25]

    2026 , eprint=

    ReasonCACHE: Teaching LLMs To Reason Without Weight Updates , author=. 2026 , eprint=

    2026

  26. [26]

    2020 , eprint=

    SuperGLUE: A Stickier Benchmark for General-Purpose Language Understanding Systems , author=. 2020 , eprint=

    2020

  27. [27]

    2022 , eprint=

    Super-NaturalInstructions: Generalization via Declarative Instructions on 1600+ NLP Tasks , author=. 2022 , eprint=

    2022

  28. [28]

    A continual learning survey: Defying forgetting in classification tasks , ISSN=

    Delange, Matthias and Aljundi, Rahaf and Masana, Marc and Parisot, Sarah and Jia, Xu and Leonardis, Ales and Slabaugh, Greg and Tuytelaars, Tinne , year=. A continual learning survey: Defying forgetting in classification tasks , ISSN=. doi:10.1109/tpami.2021.3057446 , journal=

    arXiv 2021

  29. [29]

    Neural Networks , pages=

    Continual learning: A systematic literature review , author=. Neural Networks , pages=. 2025 , publisher=

    2025

  30. [30]

    Proceedings of the 43nd International Conference on Machine Learning , year =

    Rethinking Personalization in Large Language Models at the Token Level , author=. Proceedings of the 43nd International Conference on Machine Learning , year =

  31. [31]

    2025 , eprint=

    Online Continual Learning: A Systematic Literature Review of Approaches, Challenges, and Benchmarks , author=. 2025 , eprint=

    2025

  32. [32]

    2026 , eprint=

    Memory in the Age of AI Agents , author=. 2026 , eprint=

    2026

  33. [33]

    2025 , eprint=

    From Human Memory to AI Memory: A Survey on Memory Mechanisms in the Era of LLMs , author=. 2025 , eprint=

    2025

  34. [34]

    2025 , eprint=

    A Systematic Review of Key Retrieval-Augmented Generation (RAG) Systems: Progress, Gaps, and Future Directions , author=. 2025 , eprint=

    2025

  35. [35]

    2026 , eprint=

    Continual Learning in Large Language Models: Methods, Challenges, and Opportunities , author=. 2026 , eprint=

    2026

  36. [36]

    2023 , eprint=

    Orthogonal Subspace Learning for Language Model Continual Learning , author=. 2023 , eprint=

    2023

  37. [37]

    2024 , eprint=

    Learning to Route for Dynamic Adapter Composition in Continual Learning with Language Models , author=. 2024 , eprint=

    2024

  38. [38]

    2024 , eprint=

    MixLoRA: Enhancing Large Language Models Fine-Tuning with LoRA-based Mixture of Experts , author=. 2024 , eprint=

    2024

  39. [39]

    2023 , eprint=

    QLoRA: Efficient Finetuning of Quantized LLMs , author=. 2023 , eprint=

    2023

  40. [40]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Language Ranker: A Lightweight Ranking framework for LLM Decoding , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  41. [41]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Word embeddings are steers for language models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  42. [42]

    , author=

    Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory. , author=. Psychological review , year=

  43. [43]

    Trends in cognitive sciences , volume=

    What learning systems do intelligent agents need? Complementary learning systems theory updated , author=. Trends in cognitive sciences , volume=. 2016 , publisher=

    2016

  44. [44]

    2004 , eprint=

    Clustering by compression , author=. 2004 , eprint=

    2004

  45. [45]

    Pacific-Asia conference on knowledge discovery and data mining , pages=

    Density-based clustering based on hierarchical density estimates , author=. Pacific-Asia conference on knowledge discovery and data mining , pages=. 2013 , organization=

    2013