Pith. sign in

REVIEW 3 major objections 4 minor 42 references

SoftReason makes deductive reasoning differentiable from raw pixels to the final answer, reporting 94.30 Hit@1 on knowledge-aware visual question answering.

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 09:55 UTC pith:GZNC7XJ5

load-bearing objection A genuinely new differentiable closure architecture, but the 94.30 KVQA number is unauditable until the paper says exactly how the 20-candidate entity universe is built. the 3 major comments →

arxiv 2607.20402 v1 pith:GZNC7XJ5 submitted 2026-07-22 cs.AI

SoftReason: A Fully Differentiable Neuro-Soft-Symbolic Deductive Reasoning Architecture over High-Dimensional Perceptual Data

Wael AbdAlmageed This is my paper
classification cs.AI
keywords differentiable reasoningneuro-symbolicsoft interpretation tensorimmediate-consequence operatordeductive closureknowledge graphvisual question answeringperceptual grounding
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.

SoftReason sets out to show that deductive reasoning need not begin only after perception has been converted into discrete symbols. It represents the state of what is known as a soft interpretation tensor over candidate entities and predicates, so probabilities can flow from image pixels through fact extraction, knowledge-graph evidence, and rule-based closure all the way to the answer loss. The paper's key claim is a learned differentiable version of the immediate-consequence operator—the rule step that derives new facts from existing ones—making the whole deductive loop trainable. On the KVQA benchmark under the entity-linking protocol, it reports 94.30 Hit@1 and 99.38 Recall@5, well above published baselines that rely on an external entity linker. If the claim holds, supervision from a reasoning task can shape perception itself, rather than being bolted on after a brittle symbolic handoff.

Core claim

The paper's central discovery is that deductive closure over a knowledge graph can be performed as a differentiable operator on a soft interpretation tensor F[u,p,v], whose cells approximate the probability that predicate p holds of entities (e_u,e_v). Perception first distributes image tokens over m candidate entities; known KG triples are blended in as high-confidence soft evidence via a probabilistic OR; and a learned lift of the immediate-consequence operator composes soft body-predicate mixtures across K latent channels, marginalizes over all candidate witness entities, query-conditions the head predicate, and updates the tensor by a monotone probabilistic OR. Classical Horn-chain proof

What carries the argument

The central object is the local soft interpretation tensor F∈[0,1]^{m×|P|×m}, a differentiable cell for every candidate constant pair under every predicate. The load-bearing mechanism is the learned differentiable lift of the immediate-consequence operator eTΘ: it forms soft body-predicate mixtures through trainable keys αk, βk, aggregates over all witnesses with a smoothed logical AND/OR, weighs channels by the query, gates the head predicate, and folds the proposal into the interpretation with a monotone probabilistic OR update F_{ℓ+1}=F_ℓ⊕eTΘ(F_ℓ,q). This operator's job is to let gradients pass through multi-hop deductive steps while still approximating a least-fixed-point forward-chainin

Load-bearing premise

The result stands or falls on how the m=20 candidate entities per question are chosen—the paper specifies only the budget (Supplementary, Implementation Details) and never the procedure, so if those candidates are selected with access to the dataset's named entity or the true answer, the 94.30 Hit@1 would not demonstrate reasoning from perception alone.

What would settle it

Run SoftReason with the candidate set Ex built by an independent, frozen face-identity proposer that never sees the question text or the answer labels, and compare Hit@1; if accuracy collapses, or if a control with random candidate sets still answers correctly, the perceptual reasoning claim is refuted. A second check: verify that the true answer is never among the KG triples injected at inference (ρ=0 should remove them, but the candidate list itself must also be answer-free).

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Answer-level gradients reach the perceptual encoder, so a reasoning loss can reshape visual representations rather than only score pre-extracted facts.
  • Because KG evidence is a differentiable soft injection during training and is switched off at inference, the model is trained with KG annotations yet deploys without them.
  • The fixed-point loss encourages the closure to stabilize, and classical Horn-chain reasoning is recovered as a limiting case of the learned operator.
  • On KVQA, the paper reports 94.30 Hit@1 overall and 98.20 Hit@1 on the 2-hop subset, attributing the margin to end-to-end gradient flow rather than a better entity linker.
  • The architecture is stated to be agnostic to perception modality, KG source, and rule language, needing only perceptual tokens, local symbols, and predicate-definition embeddings.

Where Pith is reading between the lines

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

  • The paper leaves implicit that the same token-plus-symbol interface should transfer to other modalities (audio, video, sensor streams); that transfer is an inference, not something KVQA validates.
  • A stress test not run in the paper: rebuild Ex with a neutral, answer-free entity proposer and re-measure Hit@1; the gap between that number and 94.30 would expose how much the differentiable closure contributes.
  • The learned channel weights could serve as a reasoning trace: reading µ_{k,p_h} per query might reveal which latent relational path was used, an interpretability avenue the paper does not pursue.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes SoftReason, a fully differentiable neuro-soft-symbolic architecture for deductive reasoning over high-dimensional perceptual inputs. The method represents the deductive state as a soft interpretation tensor over a local set of candidate constants and predicates; perception proposes probabilistic base facts, KG triples are injected as soft evidence at training time, and a learned differentiable lift of the immediate-consequence operator iteratively computes a soft closure. The resulting answer distribution is trained end-to-end with answer, fact, and fixed-point losses. The paper reports 94.30 Hit@1 on KVQA entity linking, compared with at most 62.40 for prior baselines, and argues that the end-to-end differentiability lets deductive supervision shape perceptual grounding.

Significance. If the empirical claims are substantiated, the paper's contribution is significant: it provides a concrete mechanism for keeping perception, grounding, and deductive closure in one differentiable pipeline, avoiding the discrete interface that limits classical neuro-symbolic systems. The equations are internally coherent, and the paper is honest that the closure operator is learned under supervision rather than presented as a parameter-free derivation. However, the headline result is currently not auditable: there is no code, only one run, no error bars or ablations, and the construction of the local universe is unspecified. The central empirical claim therefore rests on an unverified experimental protocol.

major comments (3)
  1. [§3.2, Eqs. (22)–(23); Supplementary, Implementation Details] The construction of the local universe E_x = {e_1,...,e_m} is never specified. The answer distribution in Eq. (23) is normalized over E_x, so Hit@1 is measured inside this candidate set. If E_x is populated using dataset-label metadata, the question's named entity, or KG lookup seeded by the gold entity, then the true answer is guaranteed to be a candidate and the 94.30 Hit@1 is a closed-vocabulary retrieval number, not an end-to-end perception-to-reasoning result. The Supplementary distinguishes the entity-linking and dataset-label-guided protocols but does not state which protocol builds E_x, nor how the m=20 candidates are proposed. This is load-bearing for Table 1 and for the §3.3 claim that 'the model reasons from perception alone.' The authors must specify the exact construction rule, state whether the gold entity can enter E_x by construction, and, if it can, report a protocol tha
  2. [§4, Table 2] Table 2 reports that the test split contains no 3-hop examples. The combined 'multi-hop' row is therefore entirely 2-hop. This does not invalidate the architecture, but it substantially weakens the claim that the learned closure operator is validated on multi-hop deduction; there is currently no evidence at depth 3. Please temper the multi-hop claim, provide 3-hop results if any exist, or re-label the evidence as 2-hop only.
  3. [§4, Table 1; Supplementary, Implementation Details] The empirical evaluation is a single run with no error bars, no seeds, no ablations, and no released code. Given the unusually large improvement (94.30 vs. 62.40), the reader cannot distinguish a genuine architectural advantage from an artifact of the candidate-entity sampling, hyperparameter choice, or evaluation protocol. At minimum, report multiple seeds with mean and variance, ablate KG injection and the closure operator (e.g., L=0, no fixed-point loss), and make the code and candidate-construction script available. Without this, the central quantitative claim is not independently reproducible.
minor comments (4)
  1. [Eq. (24)] The phrase 'F⋆[a,p_q,b] is high' is not a defined decision rule. Since the answer readout in Eq. (22)–(23) is normalized, specify the threshold or state that Eq. (24) is only an illustrative description of the reading set.
  2. [Eq. (12)] The soft-or over witnesses is written without an explicit definition. Clarify whether it is the probabilistic OR used in Eq. (10) or a different aggregation, since the later composition in Eq. (15) uses a different expression.
  3. [Abstract and §1] Minor typos: 'perceptionscomponents' in the Introduction and 'and demonstrates' in the Abstract should be corrected.
  4. [Table 1] The caption says 'closed-vocabulary accuracy' — it should be explicit that the vocabulary is the local E_x of size m=20, not the full Wikidata vocabulary, because this affects comparability with external-linker baselines.

Circularity Check

0 steps flagged

No circularity: the architecture is a supervised learning system whose 'learned lift' is fit on a training split and evaluated on a held-out test split; no load-bearing step reduces to its own inputs by construction.

full rationale

SoftReason's technical approach is an architectural construction, not a derivation of a parameter-free prediction. The core object, eT_Θ(Fℓ,q), is a learned differentiable operator trained by the answer loss L_ans (Eq. 25) against labeled examples, and the reported 94.30 Hit@1 is a held-out test result. Fitting a flexible operator to training labels and then measuring generalization is ordinary supervised empirical evaluation, not 'fitted input called prediction': the paper never claims the closure operator is derived from first principles or that a parameter-free rule set is predicted. The only potentially load-bearing premise the reader identifies—how the local universe E_x is populated—is left unspecified, but the paper nowhere defines E_x using the gold answer, Eq. (23)'s mask, or Eq. (9)'s Y_K, so no self-definitional reduction can be exhibited from the text. There are no self-citations (no prior work by the same author is invoked), no imported uniqueness theorem, and no ansatz smuggled in via citation; the claimed limiting-case recovery of Horn-chain reasoning (Eq. 18) is a stated property of the definition, not a hidden premise. The comparison to external baselines is against published results, which gives the empirical claim independent content. Thus the derivation chain is self-contained and no circular step is identifiable.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 1 invented entities

The listed learned parameters are all fitted on KVQA. The key unstated premise is the construction of the local universe Ex, which is not a parameter but determines the closed-vocabulary accuracy. The invented entity is the latent composition channel mechanism, which has no independent evidence outside the reported experiment.

free parameters (6)
  • Predicate-definition encoder ψpred and embeddings a_p
    Learned from data; the only schema-specific interface (Eq. 1).
  • Perceptual encoder fθ, grounding projection hθ, attention encoder, Wv/Ws/Wk
    Trained; fθ is a frozen ViT-B/16 per the supplementary.
  • Pre-closure bilinear heads Wp (low-rank Up,Vp) and biases bp
    Eqs. (7)-(8); the fact scorer, trained end-to-end.
  • Latent composition parameters αk, βk, γk, Wr, c_k, W_h, W_q
    Eqs. (13)-(18),(21); define the learned consequence operator.
  • KG confidence weight ρ = 0<ρ<1 during training; 0 at inference
    Controls KG evidence injection (Eq. 10); value not reported, likely chosen by hand.
  • Hyperparameters m=20, |Px|=100, K=8, L=3, λfact=1.0, λfix=0.5, rank r = m=20, |Px|=100, K=8, L=3, λfact=1.0, λfix=0.5; r unspecified
    Given in supplementary without sensitivity analysis; rank r of the low-rank factorization is not stated.
axioms (4)
  • domain assumption KG triples are trustworthy and can be used as supervision only during training; at inference KG is disabled.
    Section 3.1; this is a design choice, not proven.
  • domain assumption A local universe Ex of m=20 candidate constants contains the true answer and sufficient distractors for the closed-vocabulary evaluation.
    Supplementary implementation details; construction is unstated and load-bearing.
  • domain assumption The frozen ViT and entity grounding yield enough signal to identify the named person in KVQA images.
    Section 3.2 and implementation details; no face-recognition module is described.
  • standard math Probabilistic OR, softmax, t-norms, and monotone cell-wise updates behave as standard differentiable logic and converge when iterated L times.
    Eqs. (10),(15),(19); no formal fixed-point proof is given.
invented entities (1)
  • Latent composition channels (K=8) no independent evidence
    purpose: Learn soft body-predicate mixtures and witness aggregation for the lifted consequence operator.
    Introduced as an architectural device; validated only on one benchmark, with no external falsifiable prediction.

pith-pipeline@v1.3.0-alltime-deepseek · 10852 in / 14311 out tokens · 123617 ms · 2026-08-01T09:55:51.561598+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of SoftReason: A Fully Differentiable Neuro-Soft-Symbolic Deductive Reasoning Architecture over High-Dimensional Perceptual Data." pith.science (2026). https://pith.science/paper/GZNC7XJ5

@misc{pith2026260720402,
  author       = {Pith},
  title        = {Pith review of: SoftReason: A Fully Differentiable Neuro-Soft-Symbolic Deductive Reasoning Architecture over High-Dimensional Perceptual Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GZNC7XJ5}},
  note         = {Machine review of arXiv:2607.20402}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In many reasoning problems, the premises are not observed as discrete symbols, but must be inferred from high-dimensional inputs. Further, the predicate vocabulary, argument structure, and trusted evidence are supplied by a Knowledge Graph (KG), or rule definitions. Classical neuro-symbolic pipelines have a discrete interface between perception and deduction. We present a neuro-soft-symbolic architecture for differentiable deductive reasoning over latent perceptual facts and knowledge-provided predicates. SoftReason removes the gradient gap by representing the deductive state as a local soft interpretation tensor over candidate constants and predicates. Perception proposes probabilistic base facts, KG triples enter as high-confidence soft evidence, and every query anchor, predicate choice, and closure update remains differentiable. Our core innovation is a learned differentiable lift of the immediate-consequence operator. It uses predicate-definition embeddings and latent composition channels to form soft body-predicate mixtures, aggregate over all possible witnesses, propose query-conditioned head facts, and update the interpretation through a monotone probabilistic OR. We instantiate the framework on Knowledge-aware Visual Question Answering (KVQA), and demonstrates how SoftReason supports end-to-end perceptual grounding, KG evidence injection, and differentiable deductive closure in one trainable architecture.

Figures

Figures reproduced from arXiv: 2607.20402 by Wael AbdAlmageed.

Figure 1
Figure 1. Figure 1: SoftReason architecture, illustrating fully differentiable perception-to-reasoning. does not assume a particular perception modality, KG source, or rule language, provided input is represented as perceptual tokens, local symbols, and predicate-definition embeddings. Perceptual tokenization and entity grounding. A perceptual encoder fθ maps x to N dense tokens, and the token matrix V is defined as shown in … view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

42 extracted references · 6 canonical work pages · 4 internal anchors

  1. [1]

    International Conference on Machine Learning , pages =

    Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks , author =. International Conference on Machine Learning , pages =. 2018 , eprint =

    2018

  2. [2]

    2019 , doi =

    Shah, Sanket and Mishra, Anand and Yadati, Naganand and Talukdar, Partha Pratim , booktitle =. 2019 , doi =

    2019

  3. [3]

    2022 , eprint =

    Schwenk, Dustin and Khandelwal, Apoorv and Clark, Christopher and Marino, Kenneth and Mottaghi, Roozbeh , booktitle =. 2022 , eprint =. doi:10.48550/arXiv.2206.01718 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2206.01718 2022

  4. [4]

    CVPR , year =

    Don't just assume; look and answer: overcoming priors for visual question answering , author =. CVPR , year =. doi:10.48550/arXiv.1712.00377 , url =. 1712.00377 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.1712.00377

  5. [5]

    ACM Computing Surveys , volume =

    Survey of hallucination in natural language generation , author =. ACM Computing Surveys , volume =. 2023 , doi =

    2023

  6. [6]

    1995 , isbn =

    Foundations of Databases , author =. 1995 , isbn =

    1995

  7. [7]

    Advances in Neural Information Processing Systems , year =

    Manhaeve, Robin and Duman. Advances in Neural Information Processing Systems , year =. doi:10.48550/arXiv.1805.10872 , url =. 1805.10872 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.1805.10872

  8. [8]

    Communications of the ACM , volume =

    Wikidata: a free collaborative knowledgebase , author =. Communications of the ACM , volume =. 2014 , doi =

    2014

  9. [9]

    2017 , doi =

    Speer, Robyn and Chin, Joshua and Havasi, Catherine , booktitle =. 2017 , doi =

    2017

  10. [10]

    KRISP: Integrating Implicit and Symbolic Knowledge for Open-Domain Knowledge-Based VQA

    Marino, Kenneth and Chen, Xinlei and Parikh, Devi and Gupta, Abhinav and Rohrbach, Marcus , booktitle =. 2021 , eprint =. doi:10.48550/arXiv.2012.11014 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2012.11014 2021

  11. [11]

    2022 , eprint =

    Gui, Liangke and Wang, Borui and Huang, Qiuyuan and Hauptmann, Alexander and Bisk, Yonatan and Gao, Jianfeng , booktitle =. 2022 , eprint =. doi:10.48550/arXiv.2112.08614 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2112.08614 2022

  12. [12]

    Advances in Neural Information Processing Systems , volume =

    Scallop: From Probabilistic Deductive Databases to Scalable Differentiable Reasoning , author =. Advances in Neural Information Processing Systems , volume =. 2021 , note =

    2021

  13. [13]

    2020 , pages =

    Yang, Zhun and Ishay, Adam and Lee, Joohyung , booktitle =. 2020 , pages =. doi:10.24963/ijcai.2020/243 , url =. 2307.07700 , archivePrefix =

    Pith/arXiv arXiv 2020

  14. [14]

    ICLR , year =

    An image is worth 16x16 words: transformers for image recognition at scale , author =. ICLR , year =. doi:10.48550/arXiv.2010.11929 , url =. 2010.11929 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.2010.11929 2010

  15. [15]

    2019 , eprint =

    Deng, Jiankang and Guo, Jia and Xue, Niannan and Zafeiriou, Stefanos , booktitle =. 2019 , eprint =. doi:10.48550/arXiv.1801.07698 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.1801.07698 2019

  16. [16]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , year =

    Improving multi-hop question answering over knowledge graphs using knowledge base embeddings , author =. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , year =. doi:10.18653/v1/2020.acl-main.412 , url =

    doi:10.18653/v1/2020.acl-main.412 2020

  17. [17]

    ICLR , year =

    Graph Attention Networks , author =. ICLR , year =. doi:10.48550/arXiv.1710.10903 , url =. 1710.10903 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.1710.10903

  18. [18]

    CVPR , year =

    Bottom-up and top-down attention for image captioning and visual question answering , author =. CVPR , year =. doi:10.48550/arXiv.1707.07998 , url =. 1707.07998 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.1707.07998

  19. [19]

    Less Is More:

    Lei, Jie and Li, Linjie and Zhou, Luowei and Gan, Zhe and Berg, Tamara L and Bansal, Mohit and Liu, Jingjing , booktitle =. Less Is More:. 2021 , pages =. doi:10.1109/CVPR46437.2021.00725 , url =. 2102.06183 , archivePrefix =

    arXiv 2021

  20. [20]

    doi:10.48550/arXiv.2303.08774 , url =

    2023 , eprint =. doi:10.48550/arXiv.2303.08774 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2303.08774 2023

  21. [21]

    2018 , doi =

    Wang, Peng and Wu, Qi and Shen, Chunhua and Dick, Anthony and van den Hengel, Anton , journal =. 2018 , doi =. 1606.05433 , archivePrefix =

    Pith/arXiv arXiv 2018

  22. [22]

    Proceedings of the AAAI Conference on Artificial Intelligence , year =

    Structured two-stream attention network for video question answering , author =. Proceedings of the AAAI Conference on Artificial Intelligence , year =. doi:10.1609/aaai.v33i01.33016391 , url =

    work page doi:10.1609/aaai.v33i01.33016391

  23. [23]

    Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence , year =

    Mucko: multi-layer cross-modal knowledge reasoning for fact-based visual question answering , author =. Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence , year =. doi:10.24963/ijcai.2020/153 , url =

    work page doi:10.24963/ijcai.2020/153 2020

  24. [24]

    ISWC , year =

    Auer, S. ISWC , year =. doi:10.1007/978-3-540-76298-0_52 , url =

    doi:10.1007/978-3-540-76298-0_52

  25. [25]

    2014 , pages =

    Tandon, Niket and de Melo, Gerard and Suchanek, Fabian and Weikum, Gerhard , booktitle =. 2014 , pages =. doi:10.1145/2556195.2556245 , url =

    arXiv 2014

  26. [26]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics , pages =

    Hypergraph Transformer: Weakly-Supervised Multi-hop Reasoning for Knowledge-based Visual Question Answering , author =. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics , pages =. 2022 , doi =

    2022

  27. [27]

    CoRR , volume =

    Neural-Symbolic Computing: An Effective Methodology for Principled Integration of Machine Learning and Reasoning , author =. CoRR , volume =. 2019 , eprint =. doi:10.48550/arXiv.1905.06088 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.1905.06088 2019

  28. [28]

    CoRR , volume =

    A Semantic Loss Function for Deep Learning with Symbolic Knowledge , author =. CoRR , volume =. 2018 , eprint =. doi:10.48550/arXiv.1711.11157 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.1711.11157 2018

  29. [29]

    Artificial Intelligence , volume =

    Logic Tensor Networks , author =. Artificial Intelligence , volume =. 2022 , doi =

    2022

  30. [30]

    TensorLog: A Differentiable Deductive Database

    Cohen, William W. , journal =. 2016 , eprint =. doi:10.48550/arXiv.1605.06523 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1605.06523 2016

  31. [31]

    Advances in Neural Information Processing Systems , year =

    End-to-End Differentiable Proving , author =. Advances in Neural Information Processing Systems , year =. doi:10.48550/arXiv.1705.11040 , url =. 1705.11040 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.1705.11040

  32. [32]

    CoRR , volume =

    Learning Explanatory Rules from Noisy Data , author =. CoRR , volume =. 2018 , eprint =. doi:10.48550/arXiv.1711.04574 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.1711.04574 2018

  33. [33]

    Neural-Symbolic VQA: Disentangling Reasoning from Vision and Language Understanding

    Yi, Kexin and Wu, Jiajun and Gan, Chuang and Torralba, Antonio and Kohli, Pushmeet and Tenenbaum, Joshua B. , booktitle =. Neural-Symbolic. 2018 , eprint =. doi:10.48550/arXiv.1810.02338 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1810.02338 2018

  34. [34]

    International Conference on Learning Representations , year =

    The Neuro-Symbolic Concept Learner: Interpreting Scenes, Words, and Sentences From Natural Supervision , author =. International Conference on Learning Representations , year =. doi:10.48550/arXiv.1904.12584 , url =. 1904.12584 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.1904.12584 1904

  35. [35]

    Advances in Neural Information Processing Systems , volume =

    Translating Embeddings for Modeling Multi-relational Data , author =. Advances in Neural Information Processing Systems , volume =. 2013 , url =

    2013

  36. [36]

    2019 , eprint =

    Sun, Zhiqing and Deng, Zhi-Hong and Nie, Jian-Yun and Tang, Jian , booktitle =. 2019 , eprint =. doi:10.48550/arXiv.1902.10197 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.1902.10197 2019

  37. [37]

    International Conference on Machine Learning , year =

    Complex Embeddings for Simple Link Prediction , author =. International Conference on Machine Learning , year =. doi:10.48550/arXiv.1606.06357 , url =. 1606.06357 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.1606.06357

  38. [38]

    Improving and Diagnosing Knowledge-Based Visual Question Answering via Entity Enhanced Knowledge Injection

    Improving and Diagnosing Knowledge-Based Visual Question Answering via Entity Enhanced Knowledge Injection , author =. Proceedings of the 1st International Workshop on Multimodal Understanding for the Web and Social Media, co-located with the Web Conference 2022 , year =. doi:10.48550/arXiv.2112.06888 , url =. 2112.06888 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2112.06888 2022

  39. [39]

    and Gu, Alex and Lipkin, Benjamin and Zhang, Cedegao E

    Olausson, Theo X. and Gu, Alex and Lipkin, Benjamin and Zhang, Cedegao E. and Solar-Lezama, Armando and Tenenbaum, Joshua B. and Levy, Roger , booktitle =. 2023 , doi =

    2023

  40. [40]

    2023 , doi =

    Pan, Liangming and Albalak, Alon and Wang, Xinyi and Wang, William Yang , booktitle =. 2023 , doi =

    2023

  41. [41]

    Proceedings of the 21st International Conference on Natural Language Processing (

    Precision Empowers, Excess Distracts: Visual Question Answering With Dynamically Infused Knowledge In Language Models , author =. Proceedings of the 21st International Conference on Natural Language Processing (. 2024 , url =

    2024

  42. [42]

    Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Kenneth Marino and Mohammad Rastegari and Ali Farhadi and Roozbeh Mottaghi , title =. Conference on Computer Vision and Pattern Recognition (CVPR) , year =