Distilling Neuro-Symbolic Programs into 3D Multi-modal LLMs

Wentao Mo; Yang Liu

arxiv: 2606.01215 · v2 · pith:YIDWGMM6new · submitted 2026-05-31 · 💻 cs.CV · cs.AI· cs.CL· cs.MM

Distilling Neuro-Symbolic Programs into 3D Multi-modal LLMs

Wentao Mo , Yang Liu This is my paper

Pith reviewed 2026-06-30 11:24 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.CLcs.MM

keywords neuro-symbolic 3D reasoningmulti-modal LLMschain-of-thought distillationspatial verificationprogram tracesopen-vocabulary concepts3D grounding

0 comments

The pith

Distilling reasoning patterns from neuro-symbolic programs into 3D MLLMs unifies explicit spatial verification with open-vocabulary flexibility.

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

The paper seeks to resolve the core trade-off in 3D spatial reasoning where neuro-symbolic concept learners deliver interpretable compositional programs yet remain limited to closed vocabularies, while end-to-end 3D MLLMs manage complex language and open concepts but reason in a black-box manner. It introduces APEIRIA to transfer reasoning patterns rather than fixed concept knowledge by using natural language chain-of-thought derived from symbolic traces. A sympathetic reader would care because this could let models perform reliable stepwise checks on spatial relations in real scenes without sacrificing adaptability to new objects or instructions. The three-stage curriculum first grounds visual-geometric features, then teaches decomposition and verification from program traces via supervised fine-tuning, and finally extends the patterns to open-set cases through reinforcement learning. If the approach holds, the resulting model should match state-of-the-art MLLM performance on grounding, question answering, and captioning while retaining transparent reasoning and modular component interchangeability.

Core claim

APEIRIA is a neuro-symbolic 3D MLLM that bridges the two paradigms by distilling symbolic reasoning patterns into the model with natural language chain-of-thought. Its three-stage curriculum progressively builds capabilities through 3D perception alignment, CoT supervised fine-tuning from program traces, and CoT reinforcement learning for open-set extension. By transferring reasoning structures rather than concept-specific knowledge, the method preserves transparent reasoning and the modular interchangeability of planning and perception components, yielding performance that surpasses prior neuro-symbolic 3D methods while matching state-of-the-art 3D MLLMs on spatial reasoning benchmarks.

What carries the argument

The three-stage curriculum (perception alignment, CoT-SFT from program traces, CoT-RL) that transfers reasoning patterns rather than concept-specific knowledge into the MLLM.

If this is right

The model can process complex natural language queries and open-vocabulary concepts while retaining explicit spatial verification at each step.
Planning and perception components stay modular and interchangeable, allowing substitution without retraining the full system.
Transparent reasoning is maintained because the chain-of-thought outputs remain mappable to the original symbolic program structure.
Performance on grounding, question answering, and captioning tasks reaches or exceeds that of prior neuro-symbolic 3D methods and current 3D MLLMs.
The same distillation process extends reasoning patterns to deeply nested instructions without requiring closed-set concept limits.

Where Pith is reading between the lines

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

The same pattern-transfer approach could be tested in 2D vision or robotic planning domains where symbolic programs already exist for specific subtasks.
If modularity holds, perception modules trained on different sensor types could be swapped into the same reasoning backbone for cross-domain use.
Chain-of-thought distillation may serve as a general bridge for injecting interpretability into other large models that currently lack explicit verification.
The method implies that future work could measure how faithfully the MLLM's intermediate steps align with symbolic traces across increasing program depth.

Load-bearing premise

The three-stage curriculum successfully teaches the MLLM to perform explicit stepwise verification on open-set concepts without the verification steps becoming unreliable or non-modular.

What would settle it

A controlled test on novel 3D scenes with unseen objects where the model's generated chain-of-thought steps either fail to correctly verify a spatial relation that the original symbolic program would handle or cannot be traced back to the program's modular components.

Figures

Figures reproduced from arXiv: 2606.01215 by Wentao Mo, Yang Liu.

**Figure 1.** Figure 1: Bridging 3D MLLMs and neuro-symbolic reasoning. ➀ 3D multi-modal LLMs could handle complex natural language and open-vocabulary instructions but reason as black boxes without interpretable verification. ➁ Neuro-symbolic 3D (NS3D) methods offer transparent, step-by-step reasoning programs but are limited to closed-set concepts and require dense procedure supervision. ➂ Our method bridges them by distilling … view at source ↗

**Figure 2.** Figure 2: Progressive Curriculum-based Reasoning Distillation Framework. Left: Three-Stage Simple-to-Complex Curriculum. ➀ Perception Alignment grounds 3D visual features to textual concepts. ➁ Symbolic Reasoning Injection distills reasoning patterns from programs via CoT-SFT, providing procedural supervision for query decomposition and execution. ➂ Open-Set and Complex Reasoning Generalization extends these pattern… view at source ↗

**Figure 3.** Figure 3: Example CoT for a relate program. Net (Dai et al., 2017) and MMScan (Lyu et al., 2024) to provide ground-truth outputs for each filter operation, teaching the model basic visual-geometric semantics. Level 2 (Two-step programs): For relational reasoning, we utilize the Sr3D dataset (Achlioptas et al., 2020), which contains synthetic instructions guaranteed to be solvable in exactly two steps. Instructions i… view at source ↗

**Figure 4.** Figure 4: Example CoT revealing emerging reasoning patterns in CoT-RL stage. Object location details are omitted for brevity. the complexity of real-world instructions. 5. Conclusion We presented APEIRIA, a neuro-symbolic 3D MLLM that bridges the gap between interpretable but closed-set symbolic methods and flexible but opaque end-to-end MLLMs. Our key insight is that reasoning patterns can be distilled from symbol… view at source ↗

**Figure 5.** Figure 5: A full example of assembled thinking trace for APEIRIA. The trace includes the scene context, instruction, step-by-step thinking trace with plans and executions, and the final answer [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗

**Figure 6.** Figure 6: Modular inference enhancement examples. A.6. Limitations While APEIRIA demonstrates strong performance on 3D spatial reasoning benchmarks, several limitations merit discussion. Perception Bottleneck. As evidenced by our modularity analysis (Modularity and Scalability. in Section 4.5), the system’s overall performance remains fundamentally bounded by the accuracy of the underlying 3D perception module; when… view at source ↗

read the original abstract

Current 3D spatial reasoning methods face a fundamental trade-off: neuro-symbolic 3D (NS3D) concept learners achieve interpretable reasoning through compositional programs but are constrained to closed-set concept vocabularies and simple programs; end-to-end 3D multi-modal LLMs (3D MLLMs) could handle complex natural language and open-vocabulary concepts but suffer from black-box reasoning without explicit spatial verification. We introduce APEIRIA, a neuro-symbolic 3D MLLM to bridge two paradigms by distilling symbolic reasoning patterns into MLLMs with natural language chain-of-thought. Our three-stage curriculum progressively builds reasoning capabilities: a) 3D perception alignment grounds object visual-geometric features to the LLM, b) CoT-SFT teaches query decomposition and stepwise verification from symbolic program traces, and c) CoT-RL extends reasoning patterns to open-set concepts and deeply nested instructions. By transferring reasoning patterns rather than concept-specific knowledge, APEIRIA preserves key NS3D virtues: transparent reasoning and modular interchangeability of planning and perception components. Evaluations on grounding, question answering, and captioning show that APEIRIA surpasses prior NS3D methods and matches state-of-the-art 3D MLLMs on 3D spatial reasoning datasets, unifying symbolic methods' systematic reasoning with MLLMs' flexibility. Code is available at https://github.com/oceanflowlab/APEIRIA.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

APEIRIA's three-stage distillation of neuro-symbolic traces into 3D MLLMs is a clear new combination but the abstract supplies zero metrics or details to check if the modularity claim holds.

read the letter

The key takeaway is that this paper introduces APEIRIA, which uses a three-stage curriculum to distill neuro-symbolic programs into 3D multi-modal LLMs. The stages are perception alignment to ground features, CoT-SFT to teach decomposition from symbolic traces, and CoT-RL to handle open-set concepts. This is presented as a way to keep the transparent reasoning and modularity of neuro-symbolic methods while gaining the flexibility of LLMs for complex language and open vocabularies.

What stands out as new is the specific use of natural language chain-of-thought to transfer reasoning patterns from program traces rather than just knowledge. The paper does well in articulating the problem clearly and proposing a progressive curriculum that builds capabilities logically. The emphasis on preserving modular interchangeability of components is a strong point if it works.

Where it falls short is in the lack of any concrete results. The abstract asserts that it surpasses prior NS3D methods and matches SOTA 3D MLLMs on grounding, QA, and captioning, but gives no numbers, no dataset names, no ablations, and no discussion of how they measured the modularity or reliability of the reasoning steps. This leaves the central claim about CoT-RL maintaining explicit stepwise verification without it becoming unreliable or non-modular untested in the provided text. If the verification steps don't stay modular, the advantage over standard MLLMs disappears.

This paper is aimed at the intersection of 3D vision, multimodal models, and neuro-symbolic AI. Someone working on spatial reasoning in robotics or AR might find the distillation approach relevant. It deserves a serious referee because the idea addresses a real tension in the field and could benefit from detailed feedback on the implementation and results, assuming the full paper has the missing experimental section.

Referee Report

2 major / 0 minor

Summary. The manuscript introduces APEIRIA, a neuro-symbolic 3D MLLM that distills reasoning patterns from symbolic programs into an MLLM via a three-stage curriculum (3D perception alignment, CoT-SFT on program traces, and CoT-RL for open-set extension). It claims this preserves NS3D virtues of transparent stepwise verification and modular interchangeability of planning/perception components while matching SOTA 3D MLLM performance and surpassing prior NS3D methods on grounding, QA, and captioning tasks.

Significance. If the central claims hold with supporting evidence, the work would meaningfully bridge interpretable neuro-symbolic 3D methods and flexible end-to-end MLLMs; the public code release at the cited GitHub repository strengthens reproducibility.

major comments (2)

[Abstract] Abstract: the performance claims (surpassing prior NS3D methods and matching SOTA 3D MLLMs) are asserted without any reported metrics, dataset names, ablation studies, or error analysis, so the data-to-claim link cannot be evaluated.
[Abstract] Abstract (three-stage curriculum description): the claim that CoT-RL extends reasoning patterns to open-set concepts while preserving explicit, reliable, and modular verification steps lacks any described mechanism, ablation, or analysis showing that the steps do not collapse into non-modular pattern matching or reward hacking once the symbolic scaffold is removed.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract. We agree that strengthening the abstract with concrete evidence will improve clarity and will revise accordingly. We address each major comment below.

read point-by-point responses

Referee: [Abstract] Abstract: the performance claims (surpassing prior NS3D methods and matching SOTA 3D MLLMs) are asserted without any reported metrics, dataset names, ablation studies, or error analysis, so the data-to-claim link cannot be evaluated.

Authors: We agree the abstract would be stronger with explicit metrics. The full manuscript reports quantitative results on grounding, question answering, and captioning tasks across standard 3D spatial reasoning datasets, including direct comparisons showing APEIRIA surpassing prior NS3D methods and matching SOTA 3D MLLMs, plus ablations and error analysis in the experiments section. We will revise the abstract to include representative metrics, dataset names, and references to these supporting analyses. revision: yes
Referee: [Abstract] Abstract (three-stage curriculum description): the claim that CoT-RL extends reasoning patterns to open-set concepts while preserving explicit, reliable, and modular verification steps lacks any described mechanism, ablation, or analysis showing that the steps do not collapse into non-modular pattern matching or reward hacking once the symbolic scaffold is removed.

Authors: The mechanism is described in the methods: the CoT-RL reward is explicitly tied to verification consistency from the prior CoT-SFT stage to encourage retention of modular steps. We acknowledge the abstract does not detail this or provide supporting analysis. We will revise the abstract to briefly note the reward design and add a targeted ablation in the experiments section measuring post-RL verification step consistency to address potential collapse or reward hacking. revision: yes

Circularity Check

0 steps flagged

Empirical curriculum training shows no circularity

full rationale

The paper presents a three-stage training procedure (perception alignment, CoT-SFT from program traces, CoT-RL) evaluated on external 3D spatial reasoning datasets for grounding, QA, and captioning. No equations, fitted parameters, or derivations are shown that reduce claimed performance or preservation of NS3D virtues to quantities defined by the authors' own inputs or self-citations. The central claims rest on empirical results rather than any self-referential reduction, making the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is an applied empirical machine-learning paper; the abstract contains no mathematical derivations, fitted constants, or postulated entities that function as free parameters or axioms.

pith-pipeline@v0.9.1-grok · 5794 in / 1177 out tokens · 55267 ms · 2026-06-30T11:24:00.708601+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

17 extracted references · 9 canonical work pages · 5 internal anchors

[1]

arXiv preprint arXiv:2405.10370 , year=

Chen, S., Chen, X., Zhang, C., Li, M., Yu, G., Fei, H., Zhu, H., Fan, J., and Chen, T. Ll3da: Visual interactive in- struction tuning for omni-3d understanding reasoning and planning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 26428–26438, 2024a. Chen, S., Zhu, H., Li, M., Chen, X., Guo, P., Lei, Y ., Y...

work page arXiv
[2]

Naturally supervised 3d visual grounding with language-regularized concept learners.2024 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pp

Feng, C., Hsu, J., Liu, W., and Wu, J. Naturally supervised 3d visual grounding with language-regularized concept learners.2024 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pp. 13269–13278,

2024
[3]

Ns3d: Neuro-symbolic ground- ing of 3d objects and relations.2023 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), pp

Hsu, J., Mao, J., and Wu, J. Ns3d: Neuro-symbolic ground- ing of 3d objects and relations.2023 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), pp. 2614–2623,

2023
[4]

Chat-scene: Bridging 3d scene and large language models with object identifiers.Advances in Neural Information Processing Systems, 2024a

Huang, H., Chen, Y ., Wang, Z., Huang, R., Xu, R., Wang, T., Liu, L., Cheng, X., Zhao, Y ., Pang, J., et al. Chat-scene: Bridging 3d scene and large language models with object identifiers.Advances in Neural Information Processing Systems, 2024a. Huang, J., Yong, S., Ma, X., Linghu, X., Li, P., Wang, Y ., Li, Q., Zhu, S.-C., Jia, B., and Huang, S. An embo...

work page arXiv
[5]

Muon is Scalable for LLM Training

URL https: //kellerjordan.github.io/posts/muon/. Li, X., Liu, J., Guo, Y ., Dong, H., and Liu, Y . 3d weakly supervised visual grounding at category and instance lev- els. InProceedings of the International Conference on Robotics and Automation, 2025a. Li, Y ., Wang, Z., and Liang, W. R2g: Reasoning to ground in 3d scenes.Pattern Recognition, 168:111728, ...

work page internal anchor Pith review Pith/arXiv arXiv
[6]

Language-to-space programming for training-free 3D vi- sual grounding

Mi, B., Wang, H., Wang, T., Chen, Y ., and Pang, J. Language-to-space programming for training-free 3D vi- sual grounding. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp. 3844–3864. Association for Computational Linguis- tics,

2025
[7]

GPT-4 Technical Report

OpenAI. Gpt-4 technical report,arXiv preprint arXiv:2303.08774,

work page internal anchor Pith review Pith/arXiv arXiv
[8]

LMM-R1: Empowering 3B LMMs with Strong Reasoning Abilities Through Two-Stage Rule-Based RL

Peng, Y ., Zhang, G., Zhang, M., You, Z., Liu, J., Zhu, Q., Yang, K., Xu, X., Geng, X., and Yang, X. Lmm-r1: Em- powering 3b lmms with strong reasoning abilities through two-stage rule-based rl,arXiv preprint arXiv:2503.07536,

work page internal anchor Pith review Pith/arXiv arXiv
[9]

DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

11 Distilling Neuro-Symbolic Programs into 3D Multi-modal LLMs Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y . K., Wu, Y ., and Guo, D. Deepseek- math: Pushing the limits of mathematical reasoning in open language models,arXiv preprint arXiv:2402.03300,

work page internal anchor Pith review Pith/arXiv arXiv
[10]

Qwen3 Technical Report

Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., Zheng, C., Liu, D., Zhou, F., Huang, F., Hu, F., Ge, H., Wei, H., Lin, H., Tang, J., Yang, J., Tu, J., Zhang, J., Yang, J., Yang, J., Zhou, J., Zhou, J., Lin, J., Dang, K., Bao, K., Yang, K., Yu, L., Deng, L., Li, M., Xue, M., Li, M., Zhang, P., Wang, P., Zhu, Q...

work page internal anchor Pith review Pith/arXiv arXiv
[11]

Scene-r1: Video-grounded large language models for 3d scene reasoning without 3d annotations, arXiv preprint arXiv:2506.17545,

Yuan, Z., Jiang, S., Feng, C.-M., Zhang, Y ., Cui, S., Li, Z., and Zhao, N. Scene-r1: Video-grounded large language models for 3d scene reasoning without 3d annotations, arXiv preprint arXiv:2506.17545,

work page arXiv
[12]

Scalable ob- ject relation encoding for better 3d spatial reasoning in large language models,arXiv preprint arXiv:2603.24721,

Zhou, S., Zheng, M., Zheng, F., and Liu, Y . Scalable ob- ject relation encoding for better 3d spatial reasoning in large language models,arXiv preprint arXiv:2603.24721,

work page arXiv
[13]

Supplementary Material In this supplementary material, we provide comprehensive details to support reproducibility and further understanding of our method

12 Distilling Neuro-Symbolic Programs into 3D Multi-modal LLMs A. Supplementary Material In this supplementary material, we provide comprehensive details to support reproducibility and further understanding of our method. Section A.1 presents implementation details including detailed model architecture specifications and training protocols for all three c...

2025
[14]

Where is 13 Distilling Neuro-Symbolic Programs into 3D Multi-modal LLMs the object with ID <id> located?

and attribute annotations from MMScan (Lyu et al., 2024). Sourced from ScanNet and textual attributes extracted from captions. • Object Localization:Tasks involve outputting 3D coordinates of specified objects based on its object ID (“Where is 13 Distilling Neuro-Symbolic Programs into 3D Multi-modal LLMs the object with ID <id> located?”). We utilize gro...

2024
[15]

Think about the scene first

captions where the target is unambiguous. The total alignment dataset comprises approximately 193K instruction-response pairs. Stage 2: Symbolic Reasoning Injection via CoT-SFT.For CoT-SFT, we translate neuro-symbolic programs into iterative natural language reasoning traces. Each program step is converted into aPlansentence describing the operation, and ...

2020
[16]

{description}

and Multi3DRefer (Zhang et al., 2023), which contain natural language instructions with complex nested structures and open-vocabulary concepts. Since ground-truth execution traces are unavailable for these datasets, we rely solely on outcome supervision (whether the predicted bounding box matches the target) combined with our format reward to guide explor...

2023
[17]

find the chair to the left of the desk near the window

is left to a computer (ID 4). Now, I will formulate the response based on the identified objects. [answer] I found the vase left to the computer: Object 0: At (1.15, 6.09, 1.33), size: 0.86 x 0.99 x 1.79 Figure 5.A full example of assembled thinking trace for APEIRIA. The trace includes the scene context, instruction, step-by-step thinking trace with plan...

2025

[1] [1]

arXiv preprint arXiv:2405.10370 , year=

Chen, S., Chen, X., Zhang, C., Li, M., Yu, G., Fei, H., Zhu, H., Fan, J., and Chen, T. Ll3da: Visual interactive in- struction tuning for omni-3d understanding reasoning and planning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 26428–26438, 2024a. Chen, S., Zhu, H., Li, M., Chen, X., Guo, P., Lei, Y ., Y...

work page arXiv

[2] [2]

Naturally supervised 3d visual grounding with language-regularized concept learners.2024 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pp

Feng, C., Hsu, J., Liu, W., and Wu, J. Naturally supervised 3d visual grounding with language-regularized concept learners.2024 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pp. 13269–13278,

2024

[3] [3]

Ns3d: Neuro-symbolic ground- ing of 3d objects and relations.2023 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), pp

Hsu, J., Mao, J., and Wu, J. Ns3d: Neuro-symbolic ground- ing of 3d objects and relations.2023 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), pp. 2614–2623,

2023

[4] [4]

Chat-scene: Bridging 3d scene and large language models with object identifiers.Advances in Neural Information Processing Systems, 2024a

Huang, H., Chen, Y ., Wang, Z., Huang, R., Xu, R., Wang, T., Liu, L., Cheng, X., Zhao, Y ., Pang, J., et al. Chat-scene: Bridging 3d scene and large language models with object identifiers.Advances in Neural Information Processing Systems, 2024a. Huang, J., Yong, S., Ma, X., Linghu, X., Li, P., Wang, Y ., Li, Q., Zhu, S.-C., Jia, B., and Huang, S. An embo...

work page arXiv

[5] [5]

Muon is Scalable for LLM Training

URL https: //kellerjordan.github.io/posts/muon/. Li, X., Liu, J., Guo, Y ., Dong, H., and Liu, Y . 3d weakly supervised visual grounding at category and instance lev- els. InProceedings of the International Conference on Robotics and Automation, 2025a. Li, Y ., Wang, Z., and Liang, W. R2g: Reasoning to ground in 3d scenes.Pattern Recognition, 168:111728, ...

work page internal anchor Pith review Pith/arXiv arXiv

[6] [6]

Language-to-space programming for training-free 3D vi- sual grounding

Mi, B., Wang, H., Wang, T., Chen, Y ., and Pang, J. Language-to-space programming for training-free 3D vi- sual grounding. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp. 3844–3864. Association for Computational Linguis- tics,

2025

[7] [7]

GPT-4 Technical Report

OpenAI. Gpt-4 technical report,arXiv preprint arXiv:2303.08774,

work page internal anchor Pith review Pith/arXiv arXiv

[8] [8]

LMM-R1: Empowering 3B LMMs with Strong Reasoning Abilities Through Two-Stage Rule-Based RL

Peng, Y ., Zhang, G., Zhang, M., You, Z., Liu, J., Zhu, Q., Yang, K., Xu, X., Geng, X., and Yang, X. Lmm-r1: Em- powering 3b lmms with strong reasoning abilities through two-stage rule-based rl,arXiv preprint arXiv:2503.07536,

work page internal anchor Pith review Pith/arXiv arXiv

[9] [9]

DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

11 Distilling Neuro-Symbolic Programs into 3D Multi-modal LLMs Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y . K., Wu, Y ., and Guo, D. Deepseek- math: Pushing the limits of mathematical reasoning in open language models,arXiv preprint arXiv:2402.03300,

work page internal anchor Pith review Pith/arXiv arXiv

[10] [10]

Qwen3 Technical Report

Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., Zheng, C., Liu, D., Zhou, F., Huang, F., Hu, F., Ge, H., Wei, H., Lin, H., Tang, J., Yang, J., Tu, J., Zhang, J., Yang, J., Yang, J., Zhou, J., Zhou, J., Lin, J., Dang, K., Bao, K., Yang, K., Yu, L., Deng, L., Li, M., Xue, M., Li, M., Zhang, P., Wang, P., Zhu, Q...

work page internal anchor Pith review Pith/arXiv arXiv

[11] [11]

Scene-r1: Video-grounded large language models for 3d scene reasoning without 3d annotations, arXiv preprint arXiv:2506.17545,

Yuan, Z., Jiang, S., Feng, C.-M., Zhang, Y ., Cui, S., Li, Z., and Zhao, N. Scene-r1: Video-grounded large language models for 3d scene reasoning without 3d annotations, arXiv preprint arXiv:2506.17545,

work page arXiv

[12] [12]

Scalable ob- ject relation encoding for better 3d spatial reasoning in large language models,arXiv preprint arXiv:2603.24721,

Zhou, S., Zheng, M., Zheng, F., and Liu, Y . Scalable ob- ject relation encoding for better 3d spatial reasoning in large language models,arXiv preprint arXiv:2603.24721,

work page arXiv

[13] [13]

Supplementary Material In this supplementary material, we provide comprehensive details to support reproducibility and further understanding of our method

12 Distilling Neuro-Symbolic Programs into 3D Multi-modal LLMs A. Supplementary Material In this supplementary material, we provide comprehensive details to support reproducibility and further understanding of our method. Section A.1 presents implementation details including detailed model architecture specifications and training protocols for all three c...

2025

[14] [14]

Where is 13 Distilling Neuro-Symbolic Programs into 3D Multi-modal LLMs the object with ID <id> located?

and attribute annotations from MMScan (Lyu et al., 2024). Sourced from ScanNet and textual attributes extracted from captions. • Object Localization:Tasks involve outputting 3D coordinates of specified objects based on its object ID (“Where is 13 Distilling Neuro-Symbolic Programs into 3D Multi-modal LLMs the object with ID <id> located?”). We utilize gro...

2024

[15] [15]

Think about the scene first

captions where the target is unambiguous. The total alignment dataset comprises approximately 193K instruction-response pairs. Stage 2: Symbolic Reasoning Injection via CoT-SFT.For CoT-SFT, we translate neuro-symbolic programs into iterative natural language reasoning traces. Each program step is converted into aPlansentence describing the operation, and ...

2020

[16] [16]

{description}

and Multi3DRefer (Zhang et al., 2023), which contain natural language instructions with complex nested structures and open-vocabulary concepts. Since ground-truth execution traces are unavailable for these datasets, we rely solely on outcome supervision (whether the predicted bounding box matches the target) combined with our format reward to guide explor...

2023

[17] [17]

find the chair to the left of the desk near the window

is left to a computer (ID 4). Now, I will formulate the response based on the identified objects. [answer] I found the vase left to the computer: Object 0: At (1.15, 6.09, 1.33), size: 0.86 x 0.99 x 1.79 Figure 5.A full example of assembled thinking trace for APEIRIA. The trace includes the scene context, instruction, step-by-step thinking trace with plan...

2025