Pith. sign in

REVIEW 4 major objections 3 minor 51 references

PMMC: Prospective Multimodal Memory Compilation for Long-Term LVLM Agents

T0 review · 4 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read PMMC claims that a vision-language agent can answer long-term multimodal queries more accurately and more cheaply by compiling and execution-verifying memory-access programs before queries arrive.

desk verdict A genuinely new memory-compilation framework whose headline 3.2-point gain is plausible but not yet backed by error bars; deserves review, not desk rejection. read the letter →

arxiv 2608.00962 v1 pith:OWWUZLXH submitted 2026-08-02 cs.AI

Jingyu Sun , Yan Lin , Yuyang Xue , Yifan Wang , Zhengtao Yao , Rui Qian , Zefeng Xu , Jiachen Li
show 5 more authors
Xianyang Liu Jiancheng Pan Jingyuan Sun Syed Murtuza Baker Hongpeng Zhou
This is my paper · ORCID
classification cs.AI
keywords long-termmemoryvision-languagemodelsretrieval-augmentedgenerationmultimodalagentscompilationprospectivequestionbankbenchmarks
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that a vision-language agent should not wait for a query to decide how to access its long-term memory. At each memory-consolidation step, the agent predicts the questions it is likely to be asked, compiles each predicted question into a typed, executable memory-access program, and verifies the program by running it against the currently visible memory. Accepted question-program pairs form a question bank that routes real queries to frozen access strategies, with a retrieval-augmented fallback when no route fits or evidence is missing. The reported payoff is higher answer quality and visual-evidence recall at lower query-time token and latency cost, in exchange for extra write-time computation. This matters because long-lived agents face repeated queries over the same accumulated experiences, so planning access once and reusing it is a natural division of labor.

What carries the argument

The load-bearing mechanism is the compiled, typed memory program: a bounded sequence of operators (lexical and dense text retrieval, cross-modal image retrieval, image-text and local-context expansion, temporal expansion, raw-image materialization, and evidence-control operators) lowered from a Planner strategy under a requirement contract. The Doubter's execution-grounded readiness check makes the programs trustworthy: a program is admitted to the Question Bank only if it executes successfully over $H_{\le t}$, returns non-empty evidence, reaches every required modality and capability, and satisfies source multiplicity. The same check, run without the compiler-private answer, is applied onl

What would settle it

Construct a long-term multimodal benchmark whose queries concern events or evidence that first appear after the last consolidation boundary, so prospective questions cannot anticipate them. If PMMC's coverage on that split is near zero and its harmonized judge score is no better than its native multimodal RAG fallback, the central claim holds only inside the anticipatable-coverage regime rather than as a general memory architecture.

Watch

Extended reading notes

Core claim

PMMC's central claim is that a long-term multimodal memory should be compiled prospectively. After each consolidation boundary, a Questioner generates source-grounded prospective questions, a Planner compiles each into a bounded, typed memory program over a fixed operator kernel, and a Doubter refuses to accept a program until it has executed successfully against the visible memory and returned evidence with the required modalities and multiplicity. The resulting Question Bank stores how to access memory, not what to answer. At query time, the agent matches the incoming query, re-executes the frozen program over the currently visible memory, and answers from the materialized source evidence;

Load-bearing premise

The load-bearing premise is that future information needs can be anticipated well enough during memory consolidation that precompiled programs cover most real queries; the paper's own Table 4 caps that coverage at 82.1% at K=5, leaving the remaining queries dependent on the RAG fallback.

Editorial extensions

If this is right

  • - Re-executing a frozen program over currently visible memory means the Question Bank can retrieve memories that became visible after compilation; it is a strategy index, not a timestamped snapshot.
  • - Bounded execution-grounded refinement is the main quality lever: raising the refinement limit from $r=0$ to $r=2$ raises program success from 82.7% to 92.1% and cuts fallback from 22.4% to 12.6%, while $r=3$ adds only 0.2 HJ points for 21.1% more write-time tokens.
  • - Coverage diagnostics locate the method's limit: at $K=5$ the bank covers 82.1% of future evidence needs, and oracle candidate selection reaches HJ 65.4 versus 57.0 deployed, so routing and ranking, not just question generation, are the remaining bottleneck.
  • - Raw-image access is load-bearing for visual questions: removing it drops Mem-Gallery HJ by 4.9 points and MEMLENS by 2.3 points, even though it cuts query-time tokens by 15.6%.

Reading between the lines

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

  • The paper leaves implicit that PMMC is a bet on query repetition: the write-time overhead is recovered only when one compiled memory state serves many future queries, with the paper's own amortization showing break-even at roughly two queries against agentic baselines but about eighteen against Caption-Text RAG.
  • The gap between Cover@5 (82.1%) and Match@5 (75.2%) suggests that a learned router trained on historical query-to-bank matches could recover part of the 8.4-point oracle gap without changing program compilation; that is an extension, not a claim the paper makes.
  • The compile-then-execute pattern could generalize beyond vision-language agents: any long-lived interactive system with repeated queries over stable memory, such as coding or embodied agents, could precompile verified access protocols during idle time.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 3 minor

Summary. The paper proposes PMMC (Prospective Multimodal Memory Compilation), a framework that moves part of memory-access planning from query time to memory-consolidation time. A Questioner generates prospective questions, a Planner compiles typed multimodal memory programs, and a Doubter verifies them by execution. Accepted question-program pairs are stored in a Question Bank that is used at query time for routing and evidence retrieval, with a multimodal RAG fallback. The method is evaluated on MEMLENS and Mem-Gallery with four LVLM backbones, reporting an overall macro-average Harmonized Judge score of 49.9 versus 46.7 for the strongest baseline, first place in six of eight backbone-benchmark settings, and reductions in query-time model tokens. The paper also presents ablations of the Doubter, dynamic planning, raw-image access, self-feedback rounds, and coverage diagnostics.

Significance. If the reported gains hold, PMMC is a meaningful contribution: it reframes long-term multimodal memory as a prospective compilation problem, avoids answer caching by removing compiler-private provisional answers, and grounds verification in execution traces rather than benchmark labels. The evaluation is extensive, covering two benchmarks and four backbones, and the ablations are informative. However, the central comparative claim currently lacks uncertainty quantification, and the evaluation excludes answer-refusal and conflict-detection instances, which are core long-term-memory capabilities. The abstract also claims reduced latency, but no latency measurements are reported. These issues make the significance conditional until the statistical and scoping concerns are addressed.

major comments (4)
  1. [Empirical Results; Table 1] The primary evidence is the macro-average HJ gain from 46.7 to 49.9 and the claim of first place in six of eight settings. The paper reports no variance, confidence intervals, or significance tests. Since HJ is a per-question score in {0,0.5,1}, with N=699 per MEMLENS setting and N=1,446 per Mem-Gallery setting, setting-level differences of 1.3-2.2 points are within plausible sampling error. The macro-average over eight settings has even larger uncertainty. PMMC loses two settings and wins others by margins as small as 1.2 points, so the aggregate gain could be noise. Please provide paired bootstrap or per-setting CIs/significance tests, or temper the 'first in six of eight' and 3.2-point claims accordingly.
  2. [Datasets and Evaluation] The primary evaluation excludes 90 answer-refusal instances from MEMLENS and 81 conflict-detection plus 184 answer-refusal instances from Mem-Gallery. Refusal and conflict detection are explicitly part of these benchmarks' long-term-memory capability set. The reported macro-average therefore covers only answerable evidence-recovery queries. If the paper claims improved long-term multimodal memory in general, the excluded categories should either be reported or the claims explicitly scoped to evidence-recovery. At minimum, the abstract and conclusion should not imply full-benchmark superiority.
  3. [Abstract; Empirical Results; Figure 4] The abstract states that PMMC reduces 'query-time token and latency costs,' but no latency measurement appears anywhere in the paper. Tables 2-3 and Figure 4 report QTok and amortized model-token costs (Eq. 7), which are token counts, not wall-clock latency. Latency depends on serving, parallelism, I/O, and model size, so it cannot be inferred from token counts alone. Please either add actual latency measurements or remove the latency claim from the abstract and contributions.
  4. [Implementation Details; Online Routing] The routing thresholds alpha_route=0.60 and delta=0.05, the question budget b=8, the RRF kappa=60, and the program-length cap are hand-set, and no sensitivity analysis or selection criterion is reported. Because the fallback rate (12.6%) and the balance between compiled routes and RAG depend directly on alpha_route and delta, the reported efficiency and quality gains may be sensitive to these values. Please add a sensitivity study or justify the fixed values using a validation split.
minor comments (3)
  1. [Figure 3] The radar chart with eight axes and only representative baselines is hard to read. A per-setting difference table with intervals or a bar chart with error bars would be clearer.
  2. [Table 4] The 8.4-point gap between Oracle HJ (65.4) and deployed PMMC (57.0) is an important diagnostic, but the paper does not isolate whether the gap comes from candidate ranking, program selection, or downstream answering. A short decomposition would strengthen the analysis.
  3. [Reproducibility] The paper does not mention whether code, prompts, or Question Bank examples will be released. An artifact statement would be valuable given the multi-agent design.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PMMC's central claims are empirical comparisons on external benchmarks, and the Question Bank explicitly excludes answers and gold evidence.

full rationale

PMMC's derivation chain is self-contained. The consolidation stage is explicitly cut off from benchmark content: 'Benchmark queries, reference answers, annotated clues, and evaluator-side evidence are excluded.' The Question Bank is not an answer cache: 'Provisional answers, private source support, Planner rationales, and Doubter feedback are removed. Consequently, the Question Bank specifies how memory should be accessed but cannot directly provide an answer.' At online time the provisional answer is never used: Eq. (6) feeds the answer LVLM only the actual query, query images, and program-emitted evidence, and the readiness check Eq. (5) is 'gold-free.' The Doubter's revision signal is execution-grounded ('every PMMC revision is tied to an observed execution failure and a non-overridable readiness check') rather than label-fitted. Hyperparameters such as alpha_route=0.60 and delta=0.05 are fixed thresholds, and the coverage analysis honestly reports partial success (82.1% Cover@5, 57.0 deployed vs 65.4 oracle HJ), so the paper does not define its outcome via its inputs. The only author-overlapping citation is M3-Agent (Long et al. 2025, co-authored by Yan Lin), used as a baseline and evaluated in this paper rather than invoked as proof; it is not load-bearing. Statistical concerns about absent error bars are a correctness risk, not a circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 3 invented entities

The framework relies on several domain assumptions about coverage and the sufficiency of the program kernel. The free parameters are hyperparameters that affect routing and cost, but they are not fitted to the target result in a circular way. No new physical entities are introduced.

free parameters (6)
  • routing confidence threshold alpha_route = 0.60
    A query is routed to a compiled program only if confidence exceeds this threshold; no dev/test tuning procedure is described.
  • top-two margin delta = 0.05
    Requirement for the margin between the top two candidate routes; likely chosen heuristically.
  • prospective-question budget b = 8
    Number of candidate questions generated per compilation unit; set as a budget, not fitted.
  • refinement limit r = 2 (deployed)
    Maximum number of Planner-Doubter transitions; selected based on the analysis table showing diminishing returns, thus adjusted to balance cost and quality.
  • maximum program length = 8 operators
    Upper bound on the number of operators in a compiled program.
  • RRF kappa = 60
    Reciprocal-rank fusion constant for combining retrieval rankings.
assumptions (5)
  • domain assumption The deterministic analyzer identifies all relevant evidence opportunities (visual attributes, image-text bindings, temporal updates, etc.) from the bounded compilation unit.
    If the analyzer misses an opportunity, the Questioner cannot generate a question covering that future information need.
  • domain assumption The complementary-relation support selection (temporal proximity, lexical/entity overlap, shared images, profile links) includes sufficient evidence for any future query.
    The compilation unit Ut is built from this selection; missing evidence leads to incomplete programs.
  • domain assumption The frozen operator kernel (lexical/dense text retrieval, cross-modal retrieval, expansion, raw-image materialization) can express any evidence access strategy a future query needs.
    Programs are built only from these operators; a future query requiring a different access pattern will fail readiness and fall back.
  • domain assumption Excluding answer-refusal and conflict-detection queries from the primary evaluation does not bias the comparison.
    These categories matter for long-term memory; their exclusion affects the reported aggregate scores.
  • domain assumption GPT-5-mini as a method-blind Harmonized Judge provides a valid evaluation signal.
    The headline metric relies on a single LLM judge; no correlation with human judgments is reported.
invented entities (3)
  • Question Bank
    purpose: Stores verified prospective-question/program pairs for query-time routing and evidence retrieval.
    Internal component of the method; no external falsifiable prediction beyond the paper's own evaluation.
  • Memory program
    purpose: A typed, executable access strategy that retrieves and materializes evidence for a prospective question.
    Internal representation; constrained to a frozen operator kernel.
  • Questioner, Planner, Doubter agents
    purpose: Generate prospective questions, compile programs, and verify them through execution.
    Role-based prompts of the same LVLM backbone; they are orchestration modules, not novel physical entities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PMMC: Prospective Multimodal Memory Compilation for Long-Term LVLM Agents." pith.science (2026). https://pith.science/paper/OWWUZLXH

@misc{pith2026260800962,
  author       = {Pith},
  title        = {Pith review of: PMMC: Prospective Multimodal Memory Compilation for Long-Term LVLM Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OWWUZLXH}},
  note         = {Machine review of arXiv:2608.00962}
}
read the original abstract

Long-term memory is essential for LVLM agents to maintain consistency and integrate information across extended multimodal interactions. Existing agent memory systems, however, often reduce visual experiences into textual summaries or rely on static retrieve-then-reason pipelines, which are inefficient at query time and brittle when questions require image-text binding, temporal updates, or visual details. We propose Prospective Multimodal Memory Compilation, a framework that shifts part of the memory reasoning process from query time to memory consolidation time. Given accumulated multimodal interactions, a Questioner predicts future question candidates, a Planner compiles question-conditioned multimodal memory programs, and a Doubter verifies whether the planned evidence path can support the predicted answer. The verified question-program pairs form a structured question bank for efficient query-time routing and evidence retrieval. Experiments on multimodal long-term memory benchmarks show that our method improves answer quality and visual evidence recall while reducing query-time token and latency costs. Extensive ablations analyze the effects of self-feedback, dynamic planning, raw-image access, and question bank coverage.

Figures

Figures reproduced from arXiv: 2608.00962 by the authors.

Figure 1
Figure 1. Comparison of multimodal memory-access paradigms. Fixed RAG uses a predefined memory representation and retrieval policy, whereas agentic retrieval adapts through repeated query-time reasoning. PMMC instead compiles and verifies question-conditioned memory programs during consolidation and executes them online, retaining multimodal RAG as a fallback. piles each into a typed multimodal memory program, and a Doubter v… view at source ↗
Figure 2
Figure 2. Overview of PMMC. During consolidation, a Questioner generates source-grounded prospective questions, a Planner compiles typed memory programs, and a Doubter verifies them through execution and bounded revision. Compiler-private provisional answers and source support are removed before accepted question–program pairs enter the runtime Question Bank. At query time, PMMC routes the incoming query to a frozen program, … view at source ↗
Figure 3
Figure 3. Radar-chart view of the main results using rep [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Efficiency analysis under Qwen3.5-9B. In (a), HJ is macro-averaged over MEMLENS and Mem-Gallery, and the [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 16 canonical work pages

  1. [1]

    Asai, A.; Wu, Z.; Wang, Y.; Sil, A.; and Hajishirzi, H. 2024. Self-rag: Learning to retrieve, generate, and critique through self-reflection. In International conference on learning representations, volume 2024, 9112--9141

    work page 2024

  2. [4]

    V.; Clarke, C

    Cormack, G. V.; Clarke, C. L.; and Buettcher, S. 2009. Reciprocal rank fusion outperforms condorcet and individual rank learning methods. In Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval, 758--759

    2009

  3. [6]

    J.; and Park, J

    Jeong, S.; Baek, J.; Cho, S.; Hwang, S. J.; and Park, J. C. 2024. 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), 7036--7050

    work page 2024

  4. [7]

    F.; Gao, L.; Sun, Z.; Liu, Q.; Dwivedi-Yu, J.; Yang, Y.; Callan, J.; and Neubig, G

    Jiang, Z.; Xu, F. F.; Gao, L.; Sun, Z.; Liu, Q.; Dwivedi-Yu, J.; Yang, Y.; Callan, J.; and Neubig, G. 2023. Active retrieval augmented generation. In Proceedings of the 2023 conference on empirical methods in natural language processing, 7969--7992

    work page 2023

  5. [8]

    H.; Gonzalez, J.; Zhang, H.; and Stoica, I

    Kwon, W.; Li, Z.; Zhuang, S.; Sheng, Y.; Zheng, L.; Yu, C. H.; Gonzalez, J.; Zhang, H.; and Stoica, I. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th symposium on operating systems principles, 611--626

    2023

  6. [9]

    u ttler, H.; Lewis, M.; Yih, W.-t.; Rockt \

    Lewis, P.; Perez, E.; Piktus, A.; Petroni, F.; Karpukhin, V.; Goyal, N.; K \"u ttler, H.; Lewis, M.; Yih, W.-t.; Rockt \"a schel, T.; et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33: 9459--9474

    2020

  7. [10]

    Lewis, P.; Wu, Y.; Liu, L.; Minervini, P.; K \"u ttler, H.; Piktus, A.; Stenetorp, P.; and Riedel, S. 2021. PAQ: 65 million probably-asked questions and what you can do with them. Transactions of the Association for Computational Linguistics, 9: 1098--1115

    work page 2021

  8. [13]

    Madaan, A.; Tandon, N.; Gupta, P.; Hallinan, S.; Gao, L.; Wiegreffe, S.; Alon, U.; Dziri, N.; Prabhumoye, S.; Yang, Y.; et al. 2023. Self-refine: Iterative refinement with self-feedback. Advances in neural information processing systems, 36: 46534--46594

    work page 2023

Show all 51 references
  1. [14]

    Maharana, A.; Lee, D.-H.; Tulyakov, S.; Bansal, M.; Barbieri, F.; and Fang, Y. 2024. 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), 13851--13870

    2024

  2. [16]

    Packer, C.; Fang, V.; Patil, S.; Lin, K.; Wooders, S.; and Gonzalez, J. 2023. MemGPT: towards LLMs as operating systems. ArXiv

    2023

  3. [17]

    Papineni, K.; Roukos, S.; Ward, T.; and Zhu, W.-J. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, 311--318

    2002

  4. [18]

    S.; O'Brien, J.; Cai, C

    Park, J. S.; O'Brien, J.; Cai, C. J.; Morris, M. R.; Liang, P.; and Bernstein, M. S. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, 1--22

    2023

  5. [19]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748--8763. PmLR

    2021

  6. [20]

    Rajpurkar, P.; Zhang, J.; Lopyrev, K.; and Liang, P. 2016. Squad: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 conference on empirical methods in natural language processing, 2383--2392

    2016

  7. [22]

    Shinn, N.; Cassano, F.; Gopinath, A.; Narasimhan, K.; and Yao, S. 2023. Reflexion: Language agents with verbal reinforcement learning. Advances in neural information processing systems, 36: 8634--8652

    2023

  8. [24]

    Trivedi, H.; Balasubramanian, N.; Khot, T.; and Sabharwal, A. 2023. Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. In Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long p...

    2023

  9. [27]

    Wang, Z.; Yu, W.; Ren, X.; Zhang, J.; Zhao, Y.; Saxena, R.; Cheng, L.; Wong, G.; See, S.; Minervini, P.; et al. 2026. Mmlongbench: Benchmarking long-context vision-language models effectively and thoroughly. Advances in Neural Information Processing Systems, 38

    2026

  10. [28]

    Xu, W.; Liang, Z.; Mei, K.; Gao, H.; Tan, J.; and Zhang, Y. 2026. A-mem: Agentic memory for llm agents. Advances in Neural Information Processing Systems, 38: 17577--17604

    2026

  11. [29]

    R.; and Cao, Y

    Yao, S.; Zhao, J.; Yu, D.; Shafran, I.; Narasimhan, K. R.; and Cao, Y. 2022. React: Synergizing reasoning and acting in language models. In NeurIPS 2022 Foundation Models for Decision Making Workshop

    2022

  12. [30]

    Zhai, X.; Mustafa, B.; Kolesnikov, A.; and Beyer, L. 2023. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, 11975--11986

    2023

  13. [32]

    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=

  14. [33]

    arXiv preprint arXiv:2605.14906 , year=

    Memlens: Benchmarking multimodal long-term memory in large vision-language models , author=. arXiv preprint arXiv:2605.14906 , year=

    arXiv

  15. [34]

    arXiv preprint arXiv:2601.03515 , year=

    Mem-gallery: Benchmarking multimodal long-term conversational memory for mllm agents , author=. arXiv preprint arXiv:2601.03515 , year=

  16. [35]

    arXiv preprint arXiv:2404.18532 , year=

    Milebench: Benchmarking mllms in long context , author=. arXiv preprint arXiv:2404.18532 , year=

    arXiv

  17. [36]

    Advances in Neural Information Processing Systems , volume=

    Mmlongbench: Benchmarking long-context vision-language models effectively and thoroughly , author=. Advances in Neural Information Processing Systems , volume=

  18. [37]

    Advances in neural information processing systems , volume=

    Retrieval-augmented generation for knowledge-intensive nlp tasks , author=. Advances in neural information processing systems , volume=

  19. [38]

    Proceedings of the 36th annual acm symposium on user interface software and technology , pages=

    Generative agents: Interactive simulacra of human behavior , author=. Proceedings of the 36th annual acm symposium on user interface software and technology , pages=

  20. [39]

    , author=

    MemGPT: towards LLMs as operating systems. , author=. 2023 , howpublished=

    2023

  21. [40]

    Advances in Neural Information Processing Systems , volume=

    A-mem: Agentic memory for llm agents , author=. Advances in Neural Information Processing Systems , volume=

  22. [41]

    arXiv preprint arXiv:2504.19413 , year=

    Mem0: Building production-ready ai agents with scalable long-term memory , author=. arXiv preprint arXiv:2504.19413 , year=

    arXiv

  23. [42]

    arXiv preprint arXiv:2508.09736 , year=

    Seeing, listening, remembering, and reasoning: A multimodal agent with long-term memory , author=. arXiv preprint arXiv:2508.09736 , year=

  24. [43]

    arXiv preprint arXiv:2602.07624 , year=

    M2a: Multimodal memory agent with dual-layer hybrid memory for long-term personalized interactions , author=. arXiv preprint arXiv:2602.07624 , year=

  25. [44]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  26. [45]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Sigmoid loss for language image pre-training , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  27. [46]

    arXiv preprint arXiv:2601.04720 , year=

    Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking , author=. arXiv preprint arXiv:2601.04720 , year=

    arXiv

  28. [47]

    NeurIPS 2022 Foundation Models for Decision Making Workshop , year=

    React: Synergizing reasoning and acting in language models , author=. NeurIPS 2022 Foundation Models for Decision Making Workshop , year=

    2022

  29. [48]

    Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

    Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions , author=. Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

  30. [49]

    Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    Active retrieval augmented generation , author=. Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    2023

  31. [50]

    International conference on learning representations , volume=

    Self-rag: Learning to retrieve, generate, and critique through self-reflection , author=. International conference on learning representations , volume=

  32. [51]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    Adaptive-rag: Learning to adapt retrieval-augmented large language models through question complexity , author=. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Paper...

    2024

  33. [52]

    Transactions of the Association for Computational Linguistics , volume=

    PAQ: 65 million probably-asked questions and what you can do with them , author=. Transactions of the Association for Computational Linguistics , volume=. 2021 , publisher=

    2021

  34. [53]

    arXiv preprint arXiv:2502.10976 , year=

    QuOTE: Question-Oriented Text Embeddings , author=. arXiv preprint arXiv:2502.10976 , year=

    arXiv

  35. [54]

    arXiv preprint arXiv:1807.03100 , year=

    Robust text-to-sql generation with execution-guided decoding , author=. arXiv preprint arXiv:1807.03100 , year=

    arXiv

  36. [55]

    Advances in neural information processing systems , volume=

    Self-refine: Iterative refinement with self-feedback , author=. Advances in neural information processing systems , volume=

  37. [56]

    Advances in neural information processing systems , volume=

    Reflexion: Language agents with verbal reinforcement learning , author=. Advances in neural information processing systems , volume=

  38. [57]

    Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval , pages=

    Reciprocal rank fusion outperforms condorcet and individual rank learning methods , author=. Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval , pages=

  39. [58]

    Proceedings of the 2016 conference on empirical methods in natural language processing , pages=

    Squad: 100,000+ questions for machine comprehension of text , author=. Proceedings of the 2016 conference on empirical methods in natural language processing , pages=

    2016

  40. [59]

    Proceedings of the 40th annual meeting of the Association for Computational Linguistics , pages=

    Bleu: a method for automatic evaluation of machine translation , author=. Proceedings of the 40th annual meeting of the Association for Computational Linguistics , pages=

  41. [60]

    arXiv preprint arXiv:2506.05176 , year=

    Qwen3 embedding: Advancing text embedding and reranking through foundation models , author=. arXiv preprint arXiv:2506.05176 , year=

    arXiv

  42. [61]

    arXiv preprint arXiv:2502.14786 , year=

    Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features , author=. arXiv preprint arXiv:2502.14786 , year=

    arXiv

  43. [62]

    Proceedings of the 29th symposium on operating systems principles , pages=

    Efficient memory management for large language model serving with pagedattention , author=. Proceedings of the 29th symposium on operating systems principles , pages=

Pith tools

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