Pith. sign in

REVIEW 2 major objections 5 minor 66 references

EAR, a two-part reflection loop, claims to make LLM-agent memory retrieval both accurate and sample-efficient by combining per-query exploration with experience replay.

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 16:46 UTC pith:HGINGKBZ

load-bearing objection The paper's claims rest on a simulated citation model, but the method is clearly specified and the empirical work is substantial; it deserves review, with simulator fidelity as the first question. the 2 major comments →

arxiv 2607.17879 v1 pith:HGINGKBZ submitted 2026-07-20 cs.AI

Exploratory and Assimilating Reflection: Reflective Recall Cycle for Long-term Memory

Ganesh Senrayan , Moyuru Yamada , Ishan Jindal , Kiran Purohit This is my paper
classification cs.AI
keywords long-term memoryadaptive retrievalsample efficiencyexperience replayexploration-exploitationrerankingLLM agentsfeedback noise
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.

EAR is proposed as an adaptive memory retrieval method for LLM agents. Its central claim is that by interleaving query-specific exploration with experience replay, a lightweight reranker can be adapted from noisy interaction feedback far more sample-efficiently than by reinforcement learning on immediate rewards alone. On two long-term dialogue benchmarks, the paper reports Recall@5 gains of up to 17.9% over a static retriever and 20.6% over the RL-update baseline, and shows the adapted retriever overtakes the static baseline after about 80 training queries rather than roughly 1,400 LLM calls. The same mechanism also balances episodic and semantic memory stores, improving downstream answer quality. If this holds, long-lived agents can personalize memory retrieval from a small number of user interactions.

Core claim

EAR keeps the first-stage retriever frozen and trains a small residual adapter. Exploratory Reflection runs several UCB-style bandit rounds per query, selecting a small slate that balances relevance, diversity, and uncertainty and collecting binary usefulness labels from an external critic. Assimilating Reflection stores these experiences and replays the most query-similar ones while updating the adapter with a feedback-weighted ranking loss. The paper claims this resolves the cold-start and sample-inefficiency of immediate-reward RL: on two long-term dialogue benchmarks, Recall@5 improves by up to 17.9% over the static retriever, the static baseline is reached with 77% fewer feedback steps,

What carries the argument

The load-bearing pieces are the Explorer, a query-local multi-armed bandit whose acquisition score combines adapted query–memory similarity, a pairwise redundancy penalty, and a UCB exploration bonus; and the Experience Buffer, which stores per-query, per-memory aggregated feedback and replays the most query-similar experiences during each update. The trainable component is a residual linear adapter—two projection matrices that shift the frozen query and memory embeddings before reranking. The Explorer is what makes training data informative and the replay is what makes the global adapter learn without many interactions; together they convert a single query into a compact, reusable experienc

Load-bearing premise

The load-bearing premise is that the simulated LLM citation signal calibrated to 88/86 precision/recall faithfully represents real feedback, and that every gold memory already lies in the frozen retriever's top-K candidate set; if either fails, the reported gains may shrink.

What would settle it

Run the same 80%-train/20%-evaluate protocol with real LLM citation scores over multiple seeds and domains, comparing EAR (Explorer disabled at test time) against the static retriever. If the average Recall@5 advantage is not positive across seeds—or if it falls far below the simulated 17.9-point gap—the central claim of sample-efficient, noise-robust adaptation fails. A single real-LLM run (70.1 vs simulated 72.8) is not enough to settle this.

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

If this is right

  • If the results hold, an agent can adapt its memory retriever to a new user or domain in about 80 queries, making long-term personalization feasible without large offline datasets.
  • The Explorer can be switched off at test time, so the learned adapter alone recovers most of the gains at nearly the same latency as a static retriever; leaving it on buys additional recall at higher compute cost.
  • EAR is model-agnostic: it improves retrieval when layered on top of an existing cross-encoder reranker and across different base retrievers, though the paper reports smaller gains for larger retrievers.
  • On mixed memory stores, EAR shifts selection toward the memory type the query actually needs, and this shift improves downstream question answering on temporal queries.
  • Because the training signal is an external critic, the same loop could in principle be driven by any usefulness signal, including downstream task success, rather than citation-style labels.

Where Pith is reading between the lines

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

  • An implication the paper leaves implicit: the frozen retriever's top-K set is a hard ceiling, so any gold memory outside it is unreachable. Combining EAR with query expansion or a second candidate generator would test that boundary.
  • The simulated citation model assumes false positives are drawn independently for each candidate. Real LLM citation errors are likely correlated—an evaluator may systematically ignore certain speakers or time periods—and whether replay averaging cancels that bias is untested.
  • The paper validates against a real LLM in a single run (Recall@5 70.1 vs 72.8 simulated). Repeating the protocol across multiple seeds and domains would establish whether the simulated gains transfer.
  • The similarity-based replay sampler outperformed reward-based and random sampling in the paper's comparison. An adaptive replay sampler—reweighting experiences by recency, uncertainty, or downstream outcome—is a natural extension the paper does not explore.

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

2 major / 5 minor

Summary. The paper proposes EAR (Exploratory-Assimilating Reflection), a framework for adapting memory retrieval in LLM agents from online feedback. It consists of a residual linear adapter on query/item embeddings, an 'Explorer' that uses UCB-based bandit selection with relevance and diversity terms to iteratively build a slate of memories for a query, and an 'Assimilating Reflection' experience-replay mechanism that updates the adapter from a buffer of past query-level feedback rather than only immediate rewards. On LongMemEval and LoCoMo with Contriever/Stella/GTE retrievers and GPT-5-Mini generation, it reports absolute Recall@5 gains up to 20.6% over the RL-update baseline (Tan et al., 2025) and 17.9% over the static retriever (Contriever), 77% fewer feedback steps to reach static-retriever performance, robustness to noisy citation labels, and improved balancing of episodic versus semantic memory stores. Most experiments use a probabilistic citation simulator calibrated to precision 88 / recall 86; a single real-LLM validation run is reported in Appendix B.4.

Significance. The framework is well designed and the paper is admirably explicit about its algorithmic details: complete pseudocode, hyperparameter values, ablations (Table 3), latency analysis (Table 8), noise-robustness curves (Fig. 5), and a real-LLM spot check (Table 9) are all included. If the reported gains hold, EAR is a practical contribution: it offers a sample-efficient way to adapt memory retrieval from interaction feedback and shows that exploration plus replay can outperform myopic RL updating. The multi-store analysis (Section 5) is a nice extension. However, the empirical core relies heavily on a simulated feedback signal whose fidelity is established only weakly, and the main comparison table is internally inconsistent with the ablation/latency tables. These issues are load-bearing and need to be addressed before the central claims can be accepted.

major comments (2)
  1. [§4.4 and Appendix A.2; Table 9] All headline results in Table 2 are obtained with the probabilistic citation simulator described in §4.4 and Appendix A.2, which independently flags each gold memory with probability 0.86 and each non-gold with probability chosen to yield precision 0.88, calibrated to Tan et al. (2025). The only real-LLM validation (Appendix B.4, Table 9) is a single run: Recall@5 = 70.1 vs. simulated 72.8, with real precision 78 (not 88), no variance, and no LoCoMo result. Because the same simulated signal drives Eq. (5), Eq. (10), and Eq. (16), the mechanism's training signal and the noise-robustness claims (§4.5.3) rest on assumptions of independent citation errors and transferable precision/recall. Please report multiple runs with real LLM citations on both benchmarks and analyze sensitivity to correlated noise.
  2. [Table 2 vs. Tables 3 and 8] The RL-update baseline (Contriever, LongMemEval) is reported as Recall@5 = 59.71 in Table 2, but the identical condition appears as 54.12 in Table 3 (Adapter [Tan et al., 2025]) and Table 8 (Reranker + RL Update). The headline +20.6% improvement in §4.5.1 is computed from 80.28 − 59.71; with 54.12 the gain is 26.2%. This discrepancy is load-bearing for the central comparison and must be resolved. Additionally, Table 2 states 'averaged over 3 runs' but reports no standard deviations; please add error bars.
minor comments (5)
  1. [Table 2] Formatting error: in the LoCoMo block, the Stella row reads '79.6134.10' and the GTE row reads '77.8135.02', with missing spaces between the F1 and J columns. Please fix.
  2. [Appendix A.2] The symbol U is used both for the citation prediction space {0,...,4} and for the UCB value U_i^(t) in Eq. (5). This is confusing; rename one of them.
  3. [Eq. (1) and Limitations] The method can only adapt the ranking of candidates that appear in the frozen retriever's top-K set C (Eq. 1). If a gold memory is not in C, neither exploration nor the adapter can surface it. This structural bound is not acknowledged in the Limitations section. Please report the oracle top-K recall of the base retrievers on the two benchmarks and discuss how it bounds the maximum achievable Recall@5 for EAR and all baselines.
  4. [Limitations section] The Limitations section does not mention the dependence on the citation simulator or the single-run real-LLM check; both should be added as limitations.
  5. [Section 3.3.1] EAR returns S^(T), the last exploration slate, rather than a final exploit set based on the learned UCB statistics. Since T is small (4), the final selection still carries exploration weight; an explicit exploit round might change the results. The current design is defensible but should be justified.

Circularity Check

0 steps flagged

No significant circularity: EAR's headline results are held-out empirical comparisons against external baselines; the simulator-fidelity concern is a validity risk, not a definitional loop.

full rationale

The paper's central claims—17.9%/12.5% Recall@5 gains over the static retriever and sample-efficiency improvements over the RL baseline—are measured on held-out queries (20% split, Section 4.4) from LongMemEval and LoCoMo, against the frozen retriever and the prior RL method. The adapter is trained with a probabilistic citation simulator (Appendix A.2) whose labels are generated from gold memories, but this is a standard supervised training signal derived from the same benchmark gold labels; it does not make the held-out evaluation a restatement of the training objective, and no equation in the paper reduces a predicted quantity to a fitted parameter. The simulated-feedback setup is a legitimate correctness/fidelity concern—especially the independence assumption in Appendix A.2 and the single-run real-LLM check in Appendix B.4—but it is not circularity: even if the simulator is miscalibrated, the reported retrieval numbers are not forced to equal the simulator's precision/recall values by construction. There is no load-bearing self-citation: the only heavily cited prior method (Tan et al., 2025) is external prior work, not the present authors' own chain, and the paper does not invoke a uniqueness theorem or smuggle in an ansatz via self-citation. The internal inconsistency in the RL baseline numbers (59.71 in Table 2 vs. 54.12 in Tables 3 and 8) is a reproducibility/correctness issue, not circularity. Overall, the derivation chain is self-contained and empirically grounded; no circular step meets the evidentiary bar.

Axiom & Free-Parameter Ledger

8 free parameters · 7 axioms · 0 invented entities

The central claims rest on eight hand-set hyperparameters (three Explorer weights, UCB coefficient, T, replay strength, baseline, simulator precision/recall), of which T and the λs are tuned directly against Recall@5 on the benchmark used for the headline numbers; seven domain/technical axioms covering gold-memory availability, candidate-set coverage, test-time critic availability, unbiased feedback, replay transfer, and the nonstandard UCB heuristic; and zero invented entities — Explorer, the Experience Buffer, and the residual adapter are fully specified algorithmic components, not postulates requiring independent evidence.

free parameters (8)
  • λ_rel (relevance weight) = 0.4
    Weight on query–memory relevance in the Explorer selection score A (Eq. 7); 'chosen empirically' (Appendix A) on the same benchmarks used for the headline claims.
  • λ_div (diversity penalty) = 0.3
    Diversity penalty weight in Eq. (7); 'chosen empirically' (Appendix A).
  • λ_exp (exploration weight) = 0.3
    Weight on the UCB exploration bonus in Eq. (7); 'chosen empirically' (Appendix A).
  • α (UCB coefficient) = 0.5
    Exploration strength in Eq. (5), 'set to 0.5 to achieve a balance between exploration and exploitation' (Appendix A).
  • T (exploration rounds per query) = 4
    Number of UCB rounds per query; tuned on LongMemEval Recall@5 (Appendix B.1, Table 6) and equal to the number of extra critic calls per query.
  • reward-simulator precision/recall = p=0.88, r=0.86
    Probabilistic citation model calibrated to 'real-world LLM precision/recall reported by Tan et al. [2025]' (Section 4.4; Appendix A.2). All training rewards and the Table 2 headline numbers are produced by this simulator.
  • baseline b = 0.5
    Action-independent baseline in the feedback-weighted loss, Eq. (10); 'to give stronger negative penalty' (Appendix A.1), inherited from Tan et al.
  • λ_rep (replay strength) = 1.0
    Weight on the replay loss in Eq. (17); set in Appendix A.1.
axioms (7)
  • domain assumption Gold memories (ground-truth citation sets) are available for every training query, and simulated citations derived from them faithfully reproduce real LLM citation feedback at precision 88 / recall 86.
    Section 4.4 and Appendix A.2: the reward simulator draws TPs/FPs independently from the gold set G. All adapter training and all Table 2 numbers flow through this model; the real-LLM check (Appendix B.4) is a single run. If real LLM citation noise is correlated across candidates or domain-specific, EAR's training signal weakens.
  • domain assumption Every gold memory lies inside the frozen retriever's top-K candidate set C (Eq. 1).
    EAR, the RL baseline, and the retriever baseline all select only from C; memories outside C are unreachable, so Recall@5 is bounded by the frozen retriever's coverage. This premise is never quantified (no recall-at-K coverage analysis).
  • domain assumption LLM citation feedback is available at deployment time when the Explorer is enabled (T extra critic calls per query).
    Section 3.6 'Explorer-assisted inference' and Appendix B.3 (16.88s/query with Explorer vs 2.30s without). The headline 17.9% gain assumes a critic with usable signal exists at inference; the no-Explorer operating point (+10.4 points) is the cheaper alternative.
  • domain assumption The per-memory binary citation y_i in {−1,+1} is a noisy-but-unbiased observation of the true query–memory utility u(x, v_i).
    Formulation in Eq. (2) and the loss in Eq. (10) rely on reward correctness; systematic citation bias (e.g., LLM over-citing recent memories) would corrupt the adapter update. The paper tests noise level (Figure 5) but not bias.
  • domain assumption Replaying the top-B past experiences by query-embedding similarity improves the current update (Eq. 15).
    Appendix B.2 shows similarity-based sampling beats reward-based and random sampling empirically (72.8 vs 69.9 vs 70.3) on LongMemEval; there is no theoretical account of when the transfer holds.
  • ad hoc to paper The UCB exploration bonus in Eq. (5) is a valid uncertainty estimate for the query-local bandit.
    Eq. (5) uses a nonstandard functional form log((1+ΣN)/(1+N_i)) under the square root, with no derivation from a concentration inequality and no citation to the UCB literature. Its behavior is only validated empirically via the T=4 tuning in Appendix B.1.
  • standard math Softmax and Gumbel-top-k reparameterization (Eqs. 4, 19) give unbiased gradient estimates for the feedback-weighted ranking loss.
    Algorithm 3, line 18 and Eq. (19); standard differentiable-sampling machinery, no novelty claimed.

pith-pipeline@v1.3.0-alltime-deepseek · 5960 in / 6533 out tokens · 280744 ms · 2026-08-01T16:46:19.580312+00:00 · methodology

0 comments
read the original abstract

LLM-based autonomous agents require external memory to overcome their statelessness and limited context window for long-term interaction and dynamic knowledge reasoning. However, existing memory retrieval methods often lack adaptability and sample efficiency, and struggle to retrieve the right mixture of memories from heterogeneous stores. We propose Exploratory-Assimilating Reflection (EAR), a framework for high initial retrieval performance and sample-efficient adaptation. EAR combines two mechanisms: Exploratory Reflection, which performs iterative search to bootstrap retrieval and collect useful experiences for each query, and Assimilating Reflection, which replays these experiences from an Experience Buffer to refine a global reranker more efficiently than methods relying only on immediate rewards. Experiments show that EAR improves retrieval by up to 17.9% over the baseline retriever on two long-term dialogue benchmarks. We also show that EAR is highly sample-efficient and robust to noisy feedback.

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

66 extracted references · 7 linked inside Pith

  1. [1]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  2. [2]

    Publications Manual , year = "1983", publisher =

    1983

  3. [3]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    arXiv 1981

  4. [4]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

  5. [5]

    2021 , eprint=

    BEIR: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models , author=. 2021 , eprint=

    2021

  6. [6]

    2025 , url =

    Hit rate evaluation metric , author =. 2025 , url =

    2025

  7. [7]

    Experimental IR Meets Multilinguality, Multimodality, and Interaction (CLEF 2024) , series =

    LongEval: Longitudinal Evaluation of Model Performance at CLEF 2024 , author =. Experimental IR Meets Multilinguality, Multimodality, and Interaction (CLEF 2024) , series =. 2024 , publisher =

    2024

  8. [8]

    Marlene Gr. Team. Working Notes of. 2024 , url =

    2024

  9. [9]

    Dan Gusfield , title =. 1997

    1997

  10. [10]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  11. [11]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

  12. [12]

    In Prospect and Retrospect: Reflective Memory Management for Long-term Personalized Dialogue Agents

    Tan, Zhen and Yan, Jun and Hsu, I-Hung and Han, Rujun and Wang, Zifeng and Le, Long and Song, Yiwen and Chen, Yanfei and Palangi, Hamid and Lee, George and Iyer, Anand Rajan and Chen, Tianlong and Liu, Huan and Lee, Chen-Yu and Pfister, Tomas. In Prospect and Retrospect: Reflective Memory Management for Long-term Personalized Dialogue Agents. Proceedings ...

    2025

  13. [13]

    Neurobiology of learning and memory , volume=

    Memory systems of the brain: a brief history and current perspective , author=. Neurobiology of learning and memory , volume=. 2004 , publisher=

    2004

  14. [14]

    Nature reviews neuroscience , volume=

    Working memory: looking back and looking forward , author=. Nature reviews neuroscience , volume=. 2003 , publisher=

    2003

  15. [15]

    , author=

    The ghosts of brain states past: remembering reactivates the brain regions engaged during encoding. , author=. Psychological bulletin , volume=. 2010 , publisher=

    2010

  16. [16]

    Proceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval , pages=

    The use of MMR, diversity-based reranking for reordering documents and producing summaries , author=. Proceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval , pages=

  17. [17]

    Machine Learning , volume=

    Finite-time analysis of the multiarmed bandit problem , author=. Machine Learning , volume=. 2002 , publisher=

    2002

  18. [18]

    Machine learning , volume=

    Simple statistical gradient-following algorithms for connectionist reinforcement learning , author=. Machine learning , volume=. 1992 , publisher=

    1992

  19. [19]

    Transactions on Machine Learning Research , issn=

    Augmented Language Models: a Survey , author=. Transactions on Machine Learning Research , issn=

  20. [20]

    Retrieval-augmented generation for knowledge-intensive NLP tasks , year =

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K\". Retrieval-augmented generation for knowledge-intensive NLP tasks , year =. Proceedings of the 34th International Conference on Neural Information Processing Systems , articleno =

  21. [21]

    Yue Yu and Wei Ping and Zihan Liu and Boxin Wang and Jiaxuan You and Chao Zhang and Mohammad Shoeybi and Bryan Catanzaro , booktitle=. Rank

  22. [22]

    2023 , url=

    GPT-4 Technical Report , author=. 2023 , url=

    2023

  23. [23]

    ArXiv , year=

    The Llama 3 Herd of Models , author=. ArXiv , year=

  24. [24]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Toolformer: Language Models Can Teach Themselves to Use Tools , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

  25. [25]

    ArXiv , year=

    ReAct: Synergizing Reasoning and Acting in Language Models , author=. ArXiv , year=

  26. [26]

    2023 , publisher =

    Shinn, Noah and Cassano, Federico and Gopinath, Ashwin and Narasimhan, Karthik and Yao, Shunyu , title =. 2023 , publisher =

    2023

  27. [27]

    Robertson, Stephen and Zaragoza, Hugo , title =. Found. Trends Inf. Retr. , month = apr, pages =. 2009 , issue_date =

    2009

  28. [28]

    Transactions on Machine Learning Research , issn=

    Unsupervised Dense Information Retrieval with Contrastive Learning , author=. Transactions on Machine Learning Research , issn=. 2022 , note=

    2022

  29. [29]

    The Thirteenth International Conference on Learning Representations , year=

    LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory , author=. The Thirteenth International Conference on Learning Representations , year=

  30. [30]

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

    Evaluating very long-term conversational memory of llm agents , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  31. [31]

    The Twelfth International Conference on Learning Representations , year=

    Asai, Akari and Wu, Zeqiu and Wang, Yizhong and Sil, Avirup and Hajishirzi, Hannaneh , title=. The Twelfth International Conference on Learning Representations , year=

  32. [32]

    Adaptive- RAG : Learning to Adapt Retrieval-Augmented Large Language Models through Question Complexity

    Jeong, Soyeong and Baek, Jinheon and Cho, Sukmin and Hwang, Sung Ju and Park, Jong. Adaptive- RAG : Learning to Adapt Retrieval-Augmented Large Language Models through Question Complexity. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024

    2024

  33. [33]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    MemoryBank: Enhancing Large Language Models with Long-Term Memory , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2024 , month=

    2024

  34. [34]

    Hello Again! LLM -powered Personalized Agent for Long-term Dialogue

    Li, Hao and Yang, Chenghao and Zhang, An and Deng, Yang and Wang, Xiang and Chua, Tat-Seng. Hello Again! LLM -powered Personalized Agent for Long-term Dialogue. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025

    2025

  35. [35]

    2025 , eprint=

    Jasper and Stella: distillation of SOTA embedding models , author=. 2025 , eprint=

    2025

  36. [36]

    2023 , eprint=

    Towards General Text Embeddings with Multi-stage Contrastive Learning , author=. 2023 , eprint=

    2023

  37. [37]

    2025 , eprint=

    MIRIX: Multi-Agent Memory System for LLM-Based Agents , author=. 2025 , eprint=

    2025

  38. [38]

    2025 , eprint=

    OpenAI GPT-5 System Card , author=. 2025 , eprint=

    2025

  39. [39]

    Augmented language models: a survey

    Gr \'e goire Mialon, Roberto Dessi, Maria Lomeli, Christoforos Nalmpantis, Ramakanth Pasunuru, Roberta Raileanu, Baptiste Roziere, Timo Schick, Jane Dwivedi-Yu, Asli Celikyilmaz, Edouard Grave, Yann LeCun, and Thomas Scialom. Augmented language models: a survey. Transactions on Machine Learning Research, 2023. ISSN 2835-8856

    2023

  40. [40]

    In prospect and retrospect: Reflective memory management for long-term personalized dialogue agents

    Zhen Tan, Jun Yan, I-Hung Hsu, Rujun Han, Zifeng Wang, Long Le, Yiwen Song, Yanfei Chen, Hamid Palangi, George Lee, Anand Rajan Iyer, Tianlong Chen, Huan Liu, Chen-Yu Lee, and Tomas Pfister. In prospect and retrospect: Reflective memory management for long-term personalized dialogue agents. In Proceedings of the 63rd Annual Meeting of the Association for ...

    2025

  41. [41]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K\" u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt\" a schel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Proceedings of the 34th International Conference on Neural Information Processing Systems,...

    2020

  42. [42]

    Beir: A heterogenous benchmark for zero-shot evaluation of information retrieval models, 2021

    Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych. Beir: A heterogenous benchmark for zero-shot evaluation of information retrieval models, 2021. URL https://arxiv.org/abs/2104.08663

    Pith/arXiv arXiv 2021

  43. [43]

    Rank RAG : Unifying context ranking with retrieval-augmented generation in LLM s

    Yue Yu, Wei Ping, Zihan Liu, Boxin Wang, Jiaxuan You, Chao Zhang, Mohammad Shoeybi, and Bryan Catanzaro. Rank RAG : Unifying context ranking with retrieval-augmented generation in LLM s. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  44. [44]

    Team Gal \' a pagos Tortoise at LongEval 2024: Neural Re-Ranking and Rank Fusion for Temporal Stability

    Marlene Gr \" u ndel, Malte Weber, Johannes Franke, and Jan Heinrich Merker. Team Gal \' a pagos Tortoise at LongEval 2024: Neural Re-Ranking and Rank Fusion for Temporal Stability . In Working Notes of CLEF 2024 - Conference and Labs of the Evaluation Forum , CEUR Workshop Proceedings. CEUR-WS.org , 2024. URL https://ceur-ws.org/Vol-3740/paper-219.pdf

    2024

  45. [45]

    Longeval: Longitudinal evaluation of model performance at clef 2024

    Rabab Alkhalifa, Hsuvas Borkakoty, Romain Deveaud, Alaa El-Ebshihy, Luis Espinosa-Anke, Tobias Fink, Petra Galuščáková, Gabriela Gonzalez-Saez, Lorraine Goeuriot, David Iommi, Maria Liakata, Harish Tayyar Madabushi, Pablo Medina-Alias, Philippe Mulhem, Florina Piroi, Martin Popel, and Arkaitz Zubiaga. Longeval: Longitudinal evaluation of model performance...

    2024

  46. [46]

    Working memory: looking back and looking forward

    Alan Baddeley. Working memory: looking back and looking forward. Nature reviews neuroscience, 4 0 (10): 0 829--839, 2003

    2003

  47. [47]

    Memory systems of the brain: a brief history and current perspective

    Larry R Squire. Memory systems of the brain: a brief history and current perspective. Neurobiology of learning and memory, 82 0 (3): 0 171--177, 2004

    2004

  48. [48]

    The ghosts of brain states past: remembering reactivates the brain regions engaged during encoding

    Jared F Danker and John R Anderson. The ghosts of brain states past: remembering reactivates the brain regions engaged during encoding. Psychological bulletin, 136 0 (1): 0 87, 2010

    2010

  49. [49]

    OpenAI Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haim ing Bao, Mo Bavarian, Jeff Belgum, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenn...

    2023

  50. [50]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony S. Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aur'elien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Rozi \`e re,...

    Pith/arXiv arXiv 2024

  51. [51]

    Toolformer: Language models can teach themselves to use tools

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessi, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id=Yacmpz84TH

    2023

  52. [52]

    React: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. ArXiv, abs/2210.03629, 2022. URL https://api.semanticscholar.org/CorpusID:252762395

    Pith/arXiv arXiv 2022

  53. [53]

    Reflexion: language agents with verbal reinforcement learning

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: language agents with verbal reinforcement learning. NIPS '23, Red Hook, NY, USA, 2023. Curran Associates Inc

    2023

  54. [54]

    Memorybank: Enhancing large language models with long-term memory

    Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. Memorybank: Enhancing large language models with long-term memory. Proceedings of the AAAI Conference on Artificial Intelligence, 38 0 (17): 0 19724--19731, Mar. 2024

    2024

  55. [55]

    Hello again! LLM -powered personalized agent for long-term dialogue

    Hao Li, Chenghao Yang, An Zhang, Yang Deng, Xiang Wang, and Tat-Seng Chua. Hello again! LLM -powered personalized agent for long-term dialogue. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 5259--5276. Association for...

    2025

  56. [56]

    The probabilistic relevance framework: Bm25 and beyond

    Stephen Robertson and Hugo Zaragoza. The probabilistic relevance framework: Bm25 and beyond. Found. Trends Inf. Retr., 3 0 (4): 0 333–389, April 2009. ISSN 1554-0669. URL https://doi.org/10.1561/1500000019

    doi:10.1561/1500000019 2009

  57. [57]

    Unsupervised dense information retrieval with contrastive learning

    Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. Unsupervised dense information retrieval with contrastive learning. Transactions on Machine Learning Research, 2022. ISSN 2835-8856

    2022

  58. [58]

    Self- RAG : Learning to retrieve, generate, and critique through self-reflection

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self- RAG : Learning to retrieve, generate, and critique through self-reflection. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=hSyW5go0v8

    2024

  59. [59]

    Adaptive- RAG : Learning to adapt retrieval-augmented large language models through question complexity

    Soyeong Jeong, Jinheon Baek, Sukmin Cho, Sung Ju Hwang, and Jong Park. Adaptive- RAG : Learning to adapt retrieval-augmented large language models through question complexity. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 7036...

    2024

  60. [60]

    Jasper and stella: distillation of sota embedding models, 2025

    Dun Zhang, Jiacheng Li, Ziyang Zeng, and Fulong Wang. Jasper and stella: distillation of sota embedding models, 2025. URL https://arxiv.org/abs/2412.19048

    Pith/arXiv arXiv 2025

  61. [61]

    Towards general text embeddings with multi-stage contrastive learning, 2023

    Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. Towards general text embeddings with multi-stage contrastive learning, 2023. URL https://arxiv.org/abs/2308.03281

    Pith/arXiv arXiv 2023

  62. [62]

    Openai gpt-5 system card, 2025

    Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, Akshay Nathan, Alan Luo, Alec Helyar, Aleksander Madry, Aleksandr Efremov, Aleksandra Spyra, Alex Baker-Whitcomb, Alex Beutel, Alex Karpenko, Alex Makelov, Alex Neitz, Alex Wei, Alexandra Barr, Alexandre Kirchmeyer, Ale...

    Pith/arXiv arXiv 2025

  63. [63]

    Longmemeval: Benchmarking chat assistants on long-term interactive memory

    Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai-Wei Chang, and Dong Yu. Longmemeval: Benchmarking chat assistants on long-term interactive memory. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  64. [64]

    Evaluating very long-term conversational memory of llm agents

    Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. Evaluating very long-term conversational memory of llm agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13851--13870, 2024

    2024

  65. [65]

    Hit rate evaluation metric

    IBM . Hit rate evaluation metric. International Business Machines Corporation, 2025. URL https://www.ibm.com/docs/en/watsonx/saas?topic=metrics-hit-rate. Accessed: 2026-03-17

    2025

  66. [66]

    Mirix: Multi-agent memory system for llm-based agents, 2025

    Yu Wang and Xi Chen. Mirix: Multi-agent memory system for llm-based agents, 2025. URL https://arxiv.org/abs/2507.07957

    Pith/arXiv arXiv 2025