Pith. sign in

REVIEW 4 major objections 5 minor 125 references

Trajectory-Aware Retrieval Agents for Temporal Decision- Making

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read For decisions from temporally ordered text, standard retrieval-augmented generation discards chronology; this paper argues that fitting a latent growth curve to retrieved evidence and iteratively adding SHAP-flagged chunks yields large, con

desk verdict The LGCM-over-embeddings idea is worth a footnote, but the proof is circular and the baselines are mismatched, so the headline results are not established. read the letter →

arxiv 2607.21625 v1 pith:RQSSUUHV submitted 2026-07-07 cs.AI

Jing Wang , Jie Shen , Xing Niu This is my paper
classification cs.AI
keywords temporaldecision-makingretrieval-augmentedgenerationlatentgrowthcurvemodelSHAP-guidedrefinementLLMagenttrajectorymodelingmedicalQAfinancialprediction
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 tries to establish that when the answer depends on how information changes over time — a patient's lab trajectory, a CEO's shifting guidance, a stock's recent history — standard retrieval-augmented generation is structurally wrong because it treats retrieved chunks as an unordered bag. TLM, the proposed agent, fits a latent growth curve model to the embedding trajectory of retrieved chunks, extracts trend, turning-point, and gap signals, and then iteratively adds evidence selected by a lightweight SHAP scorer. The central empirical claim is that this closed loop beats zero-shot LLMs and standard RAG on all three tested tasks, most strikingly in medical question answering (64.2% vs 34.6%). A sympathetic reader would care because the paper offers a concrete mechanism for making retrieval 'temporal,' and because the mechanism came from classical longitudinal statistics rather than from prompt engineering.

What carries the argument

The load-bearing object is the latent growth curve model (LGCM) fit to retrieved chunk embeddings: each chunk is embedded, assigned a normalized time index, and the sequence is regressed on time to estimate an intercept (baseline), slope (rate of change), and residuals. Thresholds on those parameters produce the three signals the pipeline acts on — trend classification, turning-point detection, and gap detection. The second mechanism is the SHAP-guided refinement loop: a small query-gated attention scorer, trained jointly with the LLM, supplies leave-one-out Shapley estimates at about ten milliseconds per chunk, and the mean embedding of positive-contribution chunks seeds a learned query tha

What would settle it

On a held-out set, run the iterative refinement and record the LLM's probability of the correct label before and after each SHAP-guided addition. If a significant fraction of iterations show a decrease — especially when the scorer's φ̂_i is positive — monotonicity is empirically violated. A second check: re-fit the LGCM to chunks in shuffled order; if accuracy does not drop materially, the trajectory summary is not the source of the gains.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that chronological order in retrieved evidence is both recoverable and exploitable. The latent growth curve model decomposes chunk embeddings into a baseline, a linear rate of change, and residuals, and from that decomposition yields three operational signals: trend category, turning points, and trajectory gaps that trigger targeted re-retrieval. A jointly trained lightweight scorer lets the pipeline estimate leave-one-out SHAP values in milliseconds, so evidence refinement can be iterated without repeated LLM forward passes. The paper further asserts a monotonicity guarantee: under a scorer-calibration assumption, additive refinement never decrease

Load-bearing premise

The monotonicity guarantee needs the scorer to be calibrated, meaning a positive leave-one-out score for a chunk genuinely implies that adding that chunk (and any 'semantically similar' chunk retrieved from it) raises the correct class's probability; the paper concedes this holds only approximately.

Editorial extensions

If this is right

  • Chronological ordering becomes a first-class input to retrieval for temporal tasks, not a nuisance to be smoothed away.
  • Because refinement is additive and argued monotone, the loop can be run many times at inference without a separate utility tracker, making closed-loop retrieval practical.
  • The import of classical longitudinal statistics into LLM agents suggests a general recipe: fit structured latent-variable models to evidence embeddings to expose what is missing.
  • On financial text, even near near-chance baselines, trajectory-aware retrieval extracts economically meaningful signal — a roughly seven-point accuracy gain over zero-shot on earnings calls, and more than eight thousand dollars in additional final capital in the gap-prediction simulation.
  • The method's largest win is where chronology is most explicit (patient trajectories), suggesting temporal structure, not just more context, drives the gains.

Reading between the lines

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

  • A direct test of the trajectory claim: shuffle chunk order within each evidence set at test time and measure accuracy; if gains persist, the model is exploiting content, and if they collapse, temporal ordering is the active ingredient.
  • The linear-growth assumption is the likely ceiling; piecewise-linear or change-point growth curves could capture abrupt clinical deterioration or market regime shifts that a single slope cannot represent.
  • The calibration assumption in Proposition 1 deserves a direct audit: compare refinement iterations where SHAP-selected chunks were added against random equal-size additions; if random additions perform similarly, the refinement gain is attributable to extra context, not to attribution quality.
  • On financial tasks, the gains may partly reflect retrieving high-volatility or high-information chunks rather than true trajectory signal; an ablation replacing the query with a variance-based selection rule would separate the two.
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 / 5 minor

Summary. The paper proposes TLM, a five-stage retrieval-augmented agent for temporally structured text: hybrid retrieval; a latent growth curve model over chunk embeddings to extract trend/turning-point/gap signals; learned re-ranking and query construction; LLM next-token classification; and SHAP-guided additive evidence refinement. It claims a monotonicity guarantee (Proposition 1) under a scorer-calibration assumption, and reports large gains over zero-shot LLM and RAG baselines on a MedQA subset, earnings-call surprise prediction, and overnight stock-gap prediction. The central contribution is the combination of LGCM trajectory modeling with SHAP-guided retrieval in the LLM agent loop.

Significance. The problem is well-motivated, and the high-level idea is worth exploring: modeling retrieved evidence as a temporal trajectory and using a cheap attention scorer for leave-one-out Shapley values is a practical way to make iterative refinement feasible; the additive-only design is a sensible safeguard. If the empirical claims survived matched-baseline and ablation tests, this would be a useful contribution to temporal decision-making with LLMs. As written, however, the paper does not establish that LGCM/SHAP cause the reported gains (LoRA fine-tuning alone could), and the main theoretical statement is circular. No code or data is provided, so the numbers cannot be independently checked.

major comments (4)
  1. [§4.5, Proposition 1] Proposition 1 is not a valid derivation. The calibration assumption ('φ̂_i>0 implies non-negative marginal contribution to P(y*|·)') already assumes the conclusion for the scorer; the proof then adds 'the assumption that semantically similar chunks have non-negative marginal contribution' and concludes non-decrease, which restates the theorem. No argument connects the scorer's aggregate to the LLM softmax probability; adding chunks changes all logits and, under attention pooling, existing chunk representations. The text concedes monotonicity 'holds approximately,' making the guarantee heuristic. Remove Proposition 1 or replace it with a non-circular statement.
  2. [§5, Tables 3/5/6] All LLM baselines are zero-shot while TLM is LoRA fine-tuned; Table 3 has no RAG+LLM baseline, Tables 5–6 do not state that BM25+LLM/Embedding+LLM are fine-tuned, and §4.6 describes LoRA only for TLM. The reported gains (+29.6, +4.4, +10.8 pts) could therefore come entirely from LoRA adaptation. Add a matched fine-tuned plain-RAG baseline and ablations dropping LGCM and SHAP; otherwise the central empirical claim is unsubstantiated.
  3. [§5.1, Table 3] The Embedding baseline (15.3%) is below the 20% random chance of 5-way classification. This is implausible for a trained scorer and suggests a bug or typo. It is used to argue that semantic similarity alone is insufficient, so it must be corrected; also report confidence intervals for all accuracies.
  4. [§5.3, Table 6] The trading comparison lacks number of trades, confidence intervals, and transaction costs. The zero-shot baseline has higher win rate (80.0 vs 71.4) and Sharpe (13.5 vs 5.62); final capital depends on trade frequency and full-investment rule. The paper itself warns that Sharpe figures are unreliable due to few trades. Report trade counts and return distributions before claiming 'economically meaningful gains.'
minor comments (5)
  1. [§4.2] Define \|\Delta e\| and the 'mean residual norm' used in the turning-point and gap thresholds.
  2. [Notation] P(y*|Z) denotes both LLM probability and scorer probability (Eq. (3) vs. Proposition 1); use distinct symbols.
  3. [Table 3] The cell reads '0.1530.642'; missing separator.
  4. [§5.1] Specify how 'questions containing patient clinical trajectories' were selected; otherwise the subset is not reproducible.
  5. [§6] The stated limitation that temporal ordering is inferred from chunk position should be tested on interleaved timelines.

Circularity Check

1 steps flagged · score 6.0 of 10

The claimed monotonicity guarantee in Proposition 1 is the calibration assumption restated; the proof imports the conclusion as an extra assumption and conflates the scorer's aggregate with the LLM softmax.

  1. self definitional [Section 4.5, Proposition 1 (statement and proof)]
    "Suppose the scorer gθ is calibrated in the sense that φ̂i>0 implies zi has non-negative marginal contribution to the true posterior P(y*|·). ... Under the assumption that semantically similar chunks have non-negative marginal contribution ... adding Znew to the evidence set does not decrease the importance-weighted aggregate score for the correct class. ... we have P(y*|Z(t+1))≥P(y*|Z(t))."

    The conclusion to prove is that adding chunks semantically similar to positive-contribution evidence does not decrease P(y*|Z). The proof assumes exactly this property: "Under the assumption that semantically similar chunks have non-negative marginal contribution ... adding Znew to the evidence set does not decrease..." The only stated premise, scorer calibration on the existing chunks, does not imply the same property for newly retrieved chunks; it is simply re-entered as an assumption. Moreover, Eq. (3) defines φ̂i using the lightweight scorer's leave-one-out probabilities, while the proposition's P(y*|·) is the LLM softmax; no equation in the paper identifies the two, so calibration is asserted for a different probability than the one in the conclusion.

full rationale

The paper's main empirical pipeline (LGCM trajectory signals, SHAP-guided additive retrieval, LoRA fine-tuning) is not circular in itself, and the external datasets provide independent evaluation. The one place where a claimed derivation reduces to its own input is Proposition 1: the monotonicity guarantee is obtained by adding the target property as an extra assumption about semantically similar chunks, and the proof silently moves from the scorer's importance-weighted aggregate to the LLM's softmax probability. The abstract and Section 4.5 present this as a proven guarantee, so the circularity is material to the paper's theoretical claims. The missing matched fine-tuned RAG baselines are a genuine confound for the empirical comparison, but that is a validity concern rather than a definitional circularity and is not counted separately here. No load-bearing self-citation or uniqueness-imported-from-authors pattern appears.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities; the 'latent intercept' and 'latent slope' are ordinary least-squares coefficients, not latent variables in the structural-equation sense. The ledger's burden falls instead on hand-set thresholds, retrieval hyperparameters, and the calibration assumption that makes Proposition 1 true by construction.

free parameters (8)
  • LGCM trend classification thresholds = ||β|| < 0.05||μ||; cos(β,μ) > 0.3; cos(β,μ) < -0.3
    Hand-chosen in §4.2 to define stable/progressive/changing trajectories; no sensitivity analysis.
  • Gap detection multiplier = 1.5 · ||Δe||
    Threshold in §4.2 for flagging trajectory gaps; arbitrary.
  • Hybrid retrieval weight α = not reported
    Weights BM25 vs dense in §4.1; value never specified, yet it controls the initial evidence set.
  • Retrieval depths = kBM25=15, kemb=15, khybrid=8 (top-k sweep {3,5,8,10,12,15})
    khybrid selected on validation via sweep (§4.6); other depths fixed without justification.
  • Stock gap class thresholds = ±1%
    Discretization bounds in §5.3; no economic or statistical rationale given.
  • Event-conditioned selection threshold = 3% overnight return
    §5.3 collects only events with absolute overnight return exceeding 3%; rule is unexplained and potentially outcome-conditioned.
  • Chunk sizes and overlap = 128 words/32 overlap; 5 days/1 day
    Hyperparameters in Table 1; no sensitivity analysis.
  • SHAP feedback iterations = 1
    Table 1 sets 1 iteration, despite the paper describing an 'iterative' refinement process.
assumptions (5)
  • domain assumption Temporal ordering can be inferred from chunk position within the source document
    Acknowledged in Limitations and used throughout §4.2; however, §4.2 defines t_i from the index i in the retrieved set, and it is never stated that chunks are re-sorted by document position.
  • ad hoc to paper Scorer calibration: φ̂_i > 0 implies non-negative marginal contribution to the true posterior, and semantically similar chunks inherit non-negative contribution
    Core of the Proposition 1 proof (§4.5); the assumption is essentially the theorem's conclusion, and no evidence is given beyond 'holds approximately in practice'.
  • domain assumption Linear trajectory in BGE embedding space captures meaningful temporal structure
    §4.2 fits an OLS line through chunk embeddings; there is no validation that embedding-space progression corresponds to clinically or financially meaningful temporal change.
  • domain assumption Event-study market model with a [0,1] CAR window defines the correct earnings surprise label
    §5.2 uses CAR[0,1] > 0 as the positive label; standard, but the window choice is arbitrary and not the only defensible one.
  • ad hoc to paper Retrieval-index order imposes a time axis
    §4.2 sets t_i = (i-1)/(N-1) - 0.5 where i indexes the retrieved set; unless the set is explicitly sorted by document position (never stated), the 'trajectory' is a curve over retrieval rank, not over time.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Trajectory-Aware Retrieval Agents for Temporal Decision- Making." pith.science (2026). https://pith.science/paper/RQSSUUHV

@misc{pith2026260721625,
  author       = {Pith},
  title        = {Pith review of: Trajectory-Aware Retrieval Agents for Temporal Decision- Making},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RQSSUUHV}},
  note         = {Machine review of arXiv:2607.21625}
}
read the original abstract

We study the problem of decision-making from long-form, temporally structured text using large language model (LLM) agents. Standard retrievalaugmented generation (RAG) pipelines fragment chronological context into isolated snippets, discarding the temporal structure that is often critical for correct downstream decisions. We introduce TLM (Trajectory Language Model), a closed-loop agentic framework that iteratively refines the evidence set using SHAP-guided feedback. The key technical contribution is the latent growth curve model (LGCM) over retrieved chunk embeddings, which provides an interpretable mechanism for detecting trajectory trends, turning points, and information gaps. We show that, under a scorer-calibration assumption (which holds approximately in practice), the iterative refinement procedure is monotonically non-decreasing in the probability assigned to the correct label. Empirically, TLM is evaluated on three temporally grounded decision tasks: medical question answering, earnings call surprise prediction, and overnight stock gap prediction. TLM substantially outperforms both zero-shot LLM baselines and standard retrieval-augmented approaches on the medical task, and yields consistent, economically meaningful gains on the two financial tasks.

Figures

Figures reproduced from arXiv: 2607.21625 by the authors.

Figure 1
Figure 1. The TLM pipeline. Solid arrows indicate the forward pass; dashed red arrows indicate feedback loops. Stage 1 retrieves chunks via hybrid (BM25 + dense) search. Stage 2 fits a latent growth curve model (LGCM) to the chunk embeddings, detecting trends, turning points, and gaps; detected gaps trigger targeted re-retrieval. Stage 3 re-ranks chunks using a jointly trained scorer and constructs learned queries for additio… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

125 extracted references · 13 linked inside Pith

  1. [1]

    Attention is All you Need , url =

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N and Kaiser, ukasz and Polosukhin, Illia , booktitle =. Attention is All you Need , url =

  2. [2]

    Psychometrika , volume=

    Latent curve analysis , author=. Psychometrika , volume=. 1990 , publisher=

    1990

  3. [3]

    2006 , publisher=

    Latent Curve Models: A Structural Equation Perspective , author=. 2006 , publisher=

    2006

  4. [4]

    Contributions to the Theory of Games II , editor=

    A value for n-person games , author=. Contributions to the Theory of Games II , editor=. 1953 , publisher=

    1953

  5. [5]

    Multivariate Behavioral Research , volume=

    A latent difference score approach to growth modeling , author=. Multivariate Behavioral Research , volume=. 1987 , publisher=

    1987

  6. [6]

    2017 , publisher=

    Growth Modeling: Structural Equation and Multilevel Modeling Approaches , author=. 2017 , publisher=

    2017

  7. [7]

    Alcoholism: Clinical and Experimental Research , volume=

    Integrating person-centered and variable-centered analyses: Growth mixture modeling with latent trajectory classes , author=. Alcoholism: Clinical and Experimental Research , volume=. 2000 , publisher=

    2000

  8. [8]

    2002 , edition=

    Hierarchical Linear Models: Applications and Data Analysis Methods , author=. 2002 , edition=

    2002

Show all 125 references
  1. [9]

    2013 , edition=

    An Introduction to Latent Variable Growth Curve Modeling: Concepts, Issues, and Applications , author=. 2013 , edition=

    2013

  2. [10]

    Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics , pages=

    A multi-axis annotation scheme for event temporal relations , author=. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics , pages=

  3. [11]

    When is a liability not a liability?

    Loughran, Tim and McDonald, Bill , journal=. When is a liability not a liability?. 2011 , publisher=

    2011

  4. [12]

    Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages=

    What you say and how you say it matters: Predicting stock volatility using verbal and vocal cues , author=. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages=

  5. [13]

    Proceedings of the 11th ACM International Conference on Web Search and Data Mining , pages=

    Listening to chaotic whispers: A deep learning framework for news-oriented stock trend prediction , author=. Proceedings of the 11th ACM International Conference on Web Search and Data Mining , pages=

  6. [14]

    Advances in Neural Information Processing Systems , volume=

    A unified approach to interpreting model predictions , author=. Advances in Neural Information Processing Systems , volume=

  7. [15]

    2025 , howpublished=

    Qwen3 Technical Report , author=. 2025 , howpublished=

    2025

  8. [17]

    Advances in neural information processing systems , volume=

    Toolformer: Language models can teach themselves to use tools , author=. Advances in neural information processing systems , volume=

  9. [18]

    Proceedings of the 41st International Conference on Machine Learning , pages=

    Language agent tree search unifies reasoning, acting, and planning in language models , author=. Proceedings of the 41st International Conference on Machine Learning , pages=

  10. [19]

    Findings of ACL , pages=

    Forecasting earnings surprises from conference call transcripts , author=. Findings of ACL , pages=

  11. [20]

    Journal of financial economics , volume=

    Market efficiency, long-term returns, and behavioral finance , author=. Journal of financial economics , volume=. 1998 , publisher=

    1998

  12. [21]

    NAACL , pages=

    Reducing hallucination in structured outputs via Retrieval-Augmented Generation , author=. NAACL , pages=

  13. [22]

    ICML , pages=

    Retrieval augmented language model pre-training , author=. ICML , pages=. 2020 , organization=

    2020

  14. [23]

    ICML , year=

    Partition First, Embed Later: Laplacian-Based Feature Partitioning for Refined Embedding and Visualization of High-Dimensional Data , author=. ICML , year=

  15. [24]

    International Conference on Machine Learning (ICML) , year =

    Partition First, Embed Later: Feature Partitioning for Refined Embedding and Visualization of High-Dimensional Data , author =. International Conference on Machine Learning (ICML) , year =

  16. [25]

    Retrieval-Augmented Generation for Knowledge-Intensive

    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. Advances in Neural Information Processing Systems (NeurIPS) , year =

  17. [26]

    He, Kaiming and Gkioxari, Georgia and Doll. Mask. IEEE International Conference on Computer Vision (ICCV) , year =

  18. [27]

    European Conference on Computer Vision (ECCV) , year =

    End-to-End Object Detection with Transformers , author =. European Conference on Computer Vision (ECCV) , year =

  19. [28]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Laplacian Score for Feature Selection , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  20. [29]

    Applied and Computational Harmonic Analysis , year =

    Laplacian Eigenmaps and Spectral Techniques for Embedding and Clustering , author =. Applied and Computational Harmonic Analysis , year =

  21. [30]

    OpenReview submission:

    Anonymous , journal =. OpenReview submission:. 2024 , note =

    2024

  22. [31]

    ICML , year =

    Sebastian Borgeaud and Arthur Mensch and Jordan Hoffmann and Trevor Cai and Eliza Rutherford and Katie Millican and George Bm Van Den Driessche and Jean-Baptiste Lespiau and Bogdan Damoc and Aidan Clark and others , title =. ICML , year =

  23. [32]

    ICLR , year =

    Gautier Izacard and Edouard Grave , title =. ICLR , year =

  24. [33]

    ICLR , year =

    Urvashi Khandelwal and Omer Levy and Dan Jurafsky and Luke Zettlemoyer and Mike Lewis , title =. ICLR , year =

  25. [34]

    2023 , note =

    Jiatong Li and Yunqing Liu and Wenqi Fan and Xiao-Yong Wei and Hui Liu and Jiliang Tang and Qing Li , title =. 2023 , note =

    2023

  26. [35]

    EMNLP , year =

    Sewon Min and Julian Michael and Hannaneh Hajishirzi and Luke Zettlemoyer , title =. EMNLP , year =

  27. [36]

    Findings of EMNLP , year =

    Kurt Shuster and Spencer Poff and Moya Chen and Douwe Kiela and Jason Weston , title =. Findings of EMNLP , year =

  28. [37]

    International Conference on Learning Representations (ICLR) , year =

    Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection , author =. International Conference on Learning Representations (ICLR) , year =

  29. [38]

    A Survey on

    Fan, Wenqi and Ding, Yujuan and Ning, Liangbo and Wang, Shijie and Li, Hengyun and Yin, Dawei and Chua, Tat-Seng and Li, Qing , journal =. A Survey on. 2024 , url =

    2024

  30. [39]

    Retrieval-Augmented Generation for

    Zhao, Penghao and Zhang, Hailin and Yu, Qinhan and Wang, Zhengren and Geng, Yunteng and Fu, Fangcheng and Yang, Ling and Zhang, Wentao and Jiang, Jie and Cui, Bin , journal =. Retrieval-Augmented Generation for. 2024 , url =

    2024

  31. [40]

    arXiv preprint arXiv:2506.05176 , year =

    Qwen3 Embedding: Advancing Text Embedding and Reranking Through Training on Large-Scale High-Quality Data , author =. arXiv preprint arXiv:2506.05176 , year =

    arXiv

  32. [41]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Masked-attention Mask Transformer for Universal Image Segmentation , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  33. [42]

    ICCV , year =

    Segment Anything , author =. ICCV , year =

  34. [43]

    Proceedings of the 5th International Workshop on Semantic Evaluation (SemEval) , year =

    HeidelTime: High Quality Rule-based Extraction and Normalization of Temporal Expressions , author =. Proceedings of the 5th International Workshop on Semantic Evaluation (SemEval) , year =

  35. [44]

    ACL , year =

    Jitao Xu and Josep-Maria Crego and Jean Senellart , title =. ACL , year =

  36. [45]

    ACL , year =

    Jing Xu and Arthur Szlam and Jason Weston , title =. ACL , year =

  37. [46]

    Journal of economic literature , volume=

    Event studies in economics and finance , author=. Journal of economic literature , volume=. 1997 , publisher=

    1997

  38. [47]

    The eleventh international conference on learning representations , year=

    React: Synergizing reasoning and acting in language models , author=. The eleventh international conference on learning representations , year=

  39. [48]

    Position: Truly Self-Improving Agents Require Intrinsic Metacognitive Learning , author=

  40. [49]

    Frontiers of Computer Science , volume=

    A survey on large language model based autonomous agents , author=. Frontiers of Computer Science , volume=. 2024 , publisher=

    2024

  41. [50]

    2014 , publisher=

    Practical planning: extending the classical AI planning paradigm , author=. 2014 , publisher=

    2014

  42. [51]

    1998 , publisher=

    Reinforcement learning: An introduction , author=. 1998 , publisher=

    1998

  43. [52]

    Handbook of Mathematical Economics , volume=

    Game theory models and methods in political economy , author=. Handbook of Mathematical Economics , volume=. 1981 , publisher=

    1981

  44. [53]

    Proceedings of the 31st International Conference on Computational Linguistics , pages=

    ALYMPICS: LLM agents meet game theory , author=. Proceedings of the 31st International Conference on Computational Linguistics , pages=

  45. [54]

    Science , volume=

    Social decision-making: insights from game theory and neuroscience , author=. Science , volume=. 2007 , publisher=

    2007

  46. [55]

    Proceedings of the national academy of sciences , volume=

    Equilibrium points in n-person games , author=. Proceedings of the national academy of sciences , volume=. 1950 , publisher=

    1950

  47. [56]

    Nature communications , volume=

    Matching patients to clinical trials with large language models , author=. Nature communications , volume=. 2024 , publisher=

    2024

  48. [57]

    Science , volume=

    The neural basis of economic decision-making in the ultimatum game , author=. Science , volume=. 2003 , publisher=

    2003

  49. [58]

    Computational Linguistics , volume=

    Can large language models transform computational social science? , author=. Computational Linguistics , volume=. 2024 , publisher=

    2024

  50. [59]

    1947 , publisher=

    Theory of games and economic behavior, 2nd rev , author=. 1947 , publisher=

    1947

  51. [60]

    2012 , publisher=

    Handbook of Markov decision processes: methods and applications , author=. 2012 , publisher=

    2012

  52. [61]

    2014 , publisher=

    Markov decision processes: discrete stochastic dynamic programming , author=. 2014 , publisher=

    2014

  53. [62]

    arXiv preprint arXiv:2402.03300 , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    arXiv

  54. [63]

    arXiv preprint arXiv:1707.06347 , year=

    Proximal policy optimization algorithms , author=. arXiv preprint arXiv:1707.06347 , year=

    arXiv

  55. [65]

    AMIA Summits on Translational Science Proceedings , volume=

    A large-language model framework for relative timeline extraction from pubmed case reports , author=. AMIA Summits on Translational Science Proceedings , volume=

  56. [66]

    Metric-Fair Prompting: Treating Similar Samples Similarly , author=

  57. [67]

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

    Math-shepherd: Verify and reinforce llms step-by-step without human annotations , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  58. [68]

    arXiv preprint arXiv:1312.6114 , year=

    Auto-encoding variational bayes , author=. arXiv preprint arXiv:1312.6114 , year=

    arXiv

  59. [69]

    Journal of the American statistical Association , volume=

    Variational inference: A review for statisticians , author=. Journal of the American statistical Association , volume=. 2017 , publisher=

    2017

  60. [70]

    VIKING: Deep variational inference with stochastic projections , author=

  61. [71]

    arXiv preprint arXiv:2110.14168 , year=

    Training verifiers to solve math word problems , author=. arXiv preprint arXiv:2110.14168 , year=

    arXiv

  62. [72]

    Proceedings of the 25th international conference on Machine learning , pages=

    Apprenticeship learning using linear programming , author=. Proceedings of the 25th international conference on Machine learning , pages=

  63. [73]

    arXiv preprint arXiv:2011.00583 , year=

    An overview of multi-agent reinforcement learning from game theoretical perspective , author=. arXiv preprint arXiv:2011.00583 , year=

    2011 arXiv

  64. [74]

    Nature Reviews Cancer , volume=

    Cooperation among cancer cells: applying game theory to cancer , author=. Nature Reviews Cancer , volume=. 2019 , publisher=

    2019

  65. [75]

    Machine learning , volume=

    Q-learning , author=. Machine learning , volume=. 1992 , publisher=

    1992

  66. [76]

    Transactions on Machine Learning Research , year=

    Cognitive architectures for language agents , author=. Transactions on Machine Learning Research , year=

  67. [77]

    arXiv preprint arXiv:2507.22606 , year=

    MetaAgent: Automatically Constructing Multi-Agent Systems Based on Finite State Machines , author=. arXiv preprint arXiv:2507.22606 , year=

    arXiv

  68. [78]

    RepoAudit: An Autonomous LLM-Agent for Repository-Level Code Auditing , author=

  69. [79]

    2023 , url =

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle =. 2023 , url =

    2023

  70. [81]

    Retrieval-Augmented Generation for Knowledge-Intensive

    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. Advances in Neural Information Processing Systems , volume =. 2020 , publisher =

    2020

  71. [82]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2021 , publisher =. doi:10.1609/aaai.v35i12.17325 , url =

    2021 doi

  72. [83]

    The Annals of Statistics , volume=

    Statistical behavior and consistency of classification methods based on convex risk minimization , author=. The Annals of Statistics , volume=. 2004 , publisher=

    2004

  73. [84]

    SIGIR , pages=

    A test collection for matching patients to clinical trials , author=. SIGIR , pages=

  74. [85]

    Proceedings of the 2018 conference on empirical methods in natural language processing , pages=

    HotpotQA: A dataset for diverse, explainable multi-hop question answering , author=. Proceedings of the 2018 conference on empirical methods in natural language processing , pages=

    2018

  75. [86]

    2022 , url=

    Hu, Edward J and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle=. 2022 , url=

    2022

  76. [87]

    What disease does this patient have?

    Jin, Di and Pan, Eileen and Oufattole, Nassim and Weng, Wei-Hung and Fang, Hanyi and Szolovits, Peter , journal=. What disease does this patient have?. 2021 , publisher=

    2021

  77. [89]

    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 International Conference on Learning Representations (ICLR), 2024. URL https://arxiv.org/abs/2310.11511. arXiv:2310.11511

    2024 arXiv

  78. [90]

    Reducing hallucination in structured outputs via retrieval-augmented generation

    Orlando Ayala and Patrice Bechard. Reducing hallucination in structured outputs via retrieval-augmented generation. In NAACL, pp.\ 228--238, 2024

    2024

  79. [91]

    Latent Curve Models: A Structural Equation Perspective

    Kenneth A Bollen and Patrick J Curran. Latent Curve Models: A Structural Equation Perspective. John Wiley & Sons, 2006

    2006

  80. [92]

    Improving language models by retrieving from trillions of tokens

    Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, et al. Improving language models by retrieving from trillions of tokens. In ICML, pp.\ 2206--2240, 2022

    2022

  81. [93]

    An Introduction to Latent Variable Growth Curve Modeling: Concepts, Issues, and Applications

    Terry E Duncan, Susan C Duncan, and Lisa A Strycker. An Introduction to Latent Variable Growth Curve Modeling: Concepts, Issues, and Applications. Routledge, 2nd edition, 2013

    2013

  82. [94]

    Market efficiency, long-term returns, and behavioral finance

    Eugene F Fama. Market efficiency, long-term returns, and behavioral finance. Journal of financial economics, 49 0 (3): 0 283--306, 1998

    1998

  83. [95]

    A survey on RAG meeting LLM s: Towards retrieval-augmented large language models

    Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. A survey on RAG meeting LLM s: Towards retrieval-augmented large language models. arXiv preprint arXiv:2405.06211, 2024. URL https://arxiv.org/abs/2405.06211

    2024 arXiv

  84. [96]

    Retrieval-augmented generation for large language models: A survey

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2023

    2023 arXiv

  85. [97]

    Growth Modeling: Structural Equation and Multilevel Modeling Approaches

    Kevin J Grimm, Nilam Ram, and Ryne Estabrook. Growth Modeling: Structural Equation and Multilevel Modeling Approaches. Guilford Press, 2017

    2017

  86. [98]

    Retrieval augmented language model pre-training

    Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. Retrieval augmented language model pre-training. In ICML, pp.\ 3929--3938. PMLR, 2020

    2020

  87. [99]

    LoRA : Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA : Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=nZeVKeeFYf9

    2022

  88. [100]

    Listening to chaotic whispers: A deep learning framework for news-oriented stock trend prediction

    Ziniu Hu, Weiqing Liu, Jiang Bian, Xuanzhe Liu, and Tie-Yan Liu. Listening to chaotic whispers: A deep learning framework for news-oriented stock trend prediction. In Proceedings of the 11th ACM International Conference on Web Search and Data Mining, pp.\ 261--269, 2018

    2018

  89. [101]

    Distilling knowledge from reader to retriever for question answering

    Gautier Izacard and Edouard Grave. Distilling knowledge from reader to retriever for question answering. In ICLR, 2021

    2021

  90. [102]

    What disease does this patient have? A large-scale open domain question answering dataset from medical exams

    Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. What disease does this patient have? A large-scale open domain question answering dataset from medical exams. Applied Sciences, 11 0 (14): 0 6421, 2021

    2021

  91. [103]

    Matching patients to clinical trials with large language models

    Qiao Jin, Zifeng Wang, Charalampos S Floudas, Fangyuan Chen, Changlin Gong, Dara Bracken-Clarke, Elisabetta Xue, Yifan Yang, Jimeng Sun, and Zhiyong Lu. Matching patients to clinical trials with large language models. Nature communications, 15 0 (1): 0 9074, 2024

    2024

  92. [104]

    Forecasting earnings surprises from conference call transcripts

    Ross Koval, Nicholas Andrews, and Xifeng Yan. Forecasting earnings surprises from conference call transcripts. In Findings of ACL, pp.\ 8197--8209, 2023

    2023

  93. [105]

    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 Advances i...

    2020

  94. [106]

    When is a liability not a liability? T extual analysis, dictionaries, and 10- K s

    Tim Loughran and Bill McDonald. When is a liability not a liability? T extual analysis, dictionaries, and 10- K s. The Journal of Finance, 66 0 (1): 0 35--65, 2011

    2011

  95. [107]

    A unified approach to interpreting model predictions

    Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, volume 30, 2017

    2017

  96. [108]

    Event studies in economics and finance

    A Craig MacKinlay. Event studies in economics and finance. Journal of economic literature, 35 0 (1): 0 13--39, 1997

    1997

  97. [109]

    A latent difference score approach to growth modeling

    John J McArdle. A latent difference score approach to growth modeling. Multivariate Behavioral Research, 22 0 (4): 0 479--497, 1987

    1987

  98. [110]

    Latent curve analysis

    William Meredith and John Tisak. Latent curve analysis. Psychometrika, 55 0 (1): 0 107--122, 1990

    1990

  99. [111]

    Integrating person-centered and variable-centered analyses: Growth mixture modeling with latent trajectory classes

    Bengt Muth \'e n and Linda K Muth \'e n. Integrating person-centered and variable-centered analyses: Growth mixture modeling with latent trajectory classes. Alcoholism: Clinical and Experimental Research, 24 0 (6): 0 882--891, 2000

    2000

  100. [112]

    A multi-axis annotation scheme for event temporal relations

    Qiang Ning, Hao Wu, and Dan Roth. A multi-axis annotation scheme for event temporal relations. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, pp.\ 1318--1328, 2018

    2018

  101. [113]

    O'Brien, Carrie J

    Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST '23). AC...

    2023

  102. [114]

    What you say and how you say it matters: Predicting stock volatility using verbal and vocal cues

    Yu Qin and Yi Yang. What you say and how you say it matters: Predicting stock volatility using verbal and vocal cues. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp.\ 390--401, 2019

    2019

  103. [115]

    Hierarchical Linear Models: Applications and Data Analysis Methods

    Stephen W Raudenbush and Anthony S Bryk. Hierarchical Linear Models: Applications and Data Analysis Methods. Sage Publications, 2nd edition, 2002

    2002

  104. [116]

    Toolformer: Language models can teach themselves to use tools

    Timo Schick, Jane Dwivedi-Yu, Roberto Dess \` , Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. Advances in neural information processing systems, 36: 0 68539--68...

    2023

  105. [117]

    A value for n-person games

    Lloyd S Shapley. A value for n-person games. In Harold W Kuhn and Albert W Tucker (eds.), Contributions to the Theory of Games II, pp.\ 307--317. Princeton University Press, 1953

    1953

  106. [118]

    Retrieval augmentation reduces hallucination in conversation

    Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. Retrieval augmentation reduces hallucination in conversation. In Findings of EMNLP, pp.\ 3784--3803, 2021

    2021

  107. [119]

    Heideltime: High quality rule-based extraction and normalization of temporal expressions

    Jannik Str \"o tgen and Michael Gertz. Heideltime: High quality rule-based extraction and normalization of temporal expressions. In Proceedings of the 5th International Workshop on Semantic Evaluation (SemEval), 2010. URL https://aclanthology.org/S10-1071/

    2010

  108. [120]

    Cognitive architectures for language agents

    Theodore Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas Griffiths. Cognitive architectures for language agents. Transactions on Machine Learning Research, 2023

    2023

  109. [121]

    Qwen3 technical report

    Qwen Team. Qwen3 technical report. https://modelscope.cn/models/Qwen/Qwen3-14B, 2025

    2025

  110. [122]

    A large-language model framework for relative timeline extraction from pubmed case reports

    Jing Wang and Jeremy C Weiss. A large-language model framework for relative timeline extraction from pubmed case reports. AMIA Summits on Translational Science Proceedings, 2025: 0 598, 2025

    2025

  111. [123]

    Mimic- rnum \ 4 \ -ext-22mcts: A 22 millions-event temporal clinical time-series dataset with relative timestamp for risk prediction

    Jing Wang, Xing Niu, Juyong Kim, Jie Shen, Tong Zhang, and Jeremy C Weiss. Mimic- rnum \ 4 \ -ext-22mcts: A 22 millions-event temporal clinical time-series dataset with relative timestamp for risk prediction. arXiv preprint arXiv:2505.00827, 2025

    2025

  112. [124]

    A survey on large language model based autonomous agents

    Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. A survey on large language model based autonomous agents. Frontiers of Computer Science, 18 0 (6): 0 186345, 2024

    2024

  113. [125]

    C-pack: Packaged resources to advance general chinese embedding

    Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. C-pack: Packaged resources to advance general chinese embedding. arXiv preprint arXiv:2309.07597, 2023

    2023 arXiv

  114. [126]

    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. In International Conference on Learning Representations, 2023. URL https://arxiv.org/abs/2210.03629

    2023 arXiv

  115. [127]

    Retrieval-augmented generation for AI -generated content: A survey

    Penghao Zhao, Hailin Zhang, Qinhan Yu, Zhengren Wang, Yunteng Geng, Fangcheng Fu, Ling Yang, Wentao Zhang, Jie Jiang, and Bin Cui. Retrieval-augmented generation for AI -generated content: A survey. arXiv preprint arXiv:2402.19473, 2024. URL https://arxiv.org/abs/2402.19473

    2024 arXiv

  116. [128]

    Language agent tree search unifies reasoning, acting, and planning in language models

    Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu-Xiong Wang. Language agent tree search unifies reasoning, acting, and planning in language models. In Proceedings of the 41st International Conference on Machine Learning, pp.\ 62138--62160, 2024

    2024

  117. [129]

    Can large language models transform computational social science? Computational Linguistics, 50 0 (1): 0 237--291, 2024

    Caleb Ziems, William Held, Omar Shaikh, Jiaao Chen, Zhehao Zhang, and Diyi Yang. Can large language models transform computational social science? Computational Linguistics, 50 0 (1): 0 237--291, 2024

    2024

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.