Pith. sign in

REVIEW 4 major objections 6 minor 39 references

Recursive Vision Language Models for General Symbolic Reasoning

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read R-Qwen: a frozen 27B language model with only 12M trainable parameters learns one recursive refinement operator that solves Sudoku, maze, ARC, and crossword puzzles more accurately than far larger models.

desk verdict Solid, mostly well-run paper on recursive refinement for pretrained LLMs; the method likely works, but the headline comparisons against frontier LLMs are not protocol-matched and the HSW variance claim is borrowed, not measured. read the letter →

arxiv 2608.01534 v1 pith:HTTRPCOZ submitted 2026-08-02 cs.CV

classification cs.CV
keywords recursivereasoningsymbolicARC-AGIparameter-efficientfine-tuningLoRAself-refinementconstraintprojectiondeepsupervision
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 claims that a pretrained language model can learn to solve hard symbolic puzzles—Sudoku, maze pathfinding, ARC abstract-reasoning tasks, and crosswords—by repeatedly editing its own candidate solution instead of producing one answer in a single pass. The proposed system, R-Qwen, keeps a 27-billion-parameter backbone frozen, trains just 12 million LoRA parameters as a shared refinement operator, and reports the highest accuracies across eight benchmarks, including 88.5% on Sudoku, 93.1% on Maze, and 47.8% on ARC-AGI—above prior recursive models and above far larger frontier LLMs. The central move is to make the refinement state explicit as a serialized candidate string, pass it back through the model each step, and enforce hard constraints with a deterministic projection after every generation. A second contribution, Hierarchical Supervision Weighting (HSW), exponentially down-weights later refinement steps during training and is reported to halve training time; the paper itself notes that HSW's specific contribution to the inference-time stability is not yet isolated. If the results hold, iterative self-correction and pretrained language priors are complementary, and recursive depth can substitute for parameter scale.

What carries the argument

The load-bearing mechanism is the recursive candidate-refinement loop with constraint projection. The state is explicit—a serialized candidate string $\hat{y}^{(t)}$ that the model re-encodes every step—with update $\hat{y}^{(t)} = \Pi_x(\mathrm{Decode}(P(x, c, \hat{y}^{(t-1)}, t, N)))$. The projection $\Pi_x$ enforces hard constraints by clamping the given positions and replacing invalid or missing tokens with the previous candidate's value, so every intermediate state stays feasible. Around this loop, two training devices carry the argument: deep supervision, which teacher-forces the full target at every step so the model learns to map any partially-correct candidate to the solution; and H

What would settle it

Re-run the published R-Qwen recipe with uniform step weights ($\lambda = 1$) on Sudoku-Extreme and ARC-AGI and compare convergence time, per-step gradient variance, and final exact-match accuracy against the HSW version; the paper itself states that isolating HSW's effect requires exactly this controlled ablation against uniform deep supervision and lists it as future work. Independently re-running the full training setup (frozen Qwen3.5-27B, rank-16 LoRA, stated curriculum) from released code would confirm whether the headline numbers—88.5% Sudoku, 93.1% Maze, 47.8% ARC-AGI—reproduce.

Watch

Extended reading notes

Core claim

At its center is the claim that recursive candidate refinement transfers to a pretrained autoregressive backbone without architectural change. Every puzzle becomes a serialized constraint-satisfaction instance—initial string, optional context, target solution, and a clamped set of positions that must stay fixed. One refinement operator (frozen Qwen plus LoRA) is unrolled for $T$ steps: prompted with the instance and current candidate, it decodes an improved candidate, and a task-specific projection $\Pi_x$ clamps fixed cells, discards invalid tokens, and falls back to the previous candidate. Every step is supervised by the full target, with losses exponentially down-weighted (HSW) to favor e

Load-bearing premise

The load-bearing premise is that exponentially weighting early refinement steps is what speeds up training—a variance-reduction benefit this paper takes from a prior method without measuring it here—so if plain uniform weighting trains just as fast and accurately, the optimization story gives way even though the benchmark numbers could survive.

Editorial extensions

If this is right

  • One learned refinement operator transfers across Sudoku, maze, ARC, and crossword puzzles: only the serialization, alphabet, and constraint projection change, not the training algorithm or loss.
  • Recursive depth substitutes for parameter scale: 12M trainable parameters on a frozen 27B backbone beat prior recursive models and far larger frontier LLMs on the reported benchmarks.
  • Recursion-aware training is necessary: applying recursive refinement to a pretrained model without recursive training (the paper's first experiment) does not improve accuracy, consistent with earlier evidence that LLMs cannot self-correct without supervision.
  • Evaluating beyond the training horizon stays stable and continues to improve for R-Qwen, whereas TRM, HRM, and LoopFormer fluctuate after roughly ten to sixteen steps; the paper flags that HSW's role in this stability is not yet isolated.
  • HSW roughly halves training time (2.04x for R-Qwen, 2.09x for TRM) while improving or preserving exact accuracy in the reported runs.

Reading between the lines

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

  • If the benchmark numbers reproduce, the cheapest known route to stronger constrained reasoning may be to bolt a deterministic constraint projection onto a frozen LLM and deep-supervise a recursion curriculum via LoRA—no architectural change and no reinforcement learning—which could extend to code repair, theorem proving, or planning with hard legality constraints.
  • The design gives the projection responsibility for feasibility, so an ablation that removes the fallback behavior (keeping only clamping) would reveal how much of the gain is learned correction versus the projection doing the constraint-satisfaction work.
  • Because the paper states R-Qwen is orthogonal to vision-language synergy, a concrete untested next step is grafting a visual co-reasoning module onto the recursive operator, which could push ARC-style results beyond the reported 47.8%.
  • The paper leaves the uniform-weighting comparison ($\lambda = 1$) to future work; if that null run matches HSW's accuracy and speed, exponential weighting should be read as a training-efficiency device rather than the source of the reasoning gains.
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 / 6 minor

Summary. The paper introduces R-Qwen, a recursive candidate-refinement framework that fine-tunes a frozen pretrained Qwen3.5-27B backbone with LoRA. A shared refinement operator receives the original problem, the current serialized candidate, and a step index; decodes an improved candidate; applies a task-specific constraint projection; and feeds the result back. All tasks (Sudoku, Maze, ARC-1/2, ARC-AGI, BARC, Re-ARC, CrossWordBench) are cast as serialized constraint satisfaction. Training uses deep supervision at every step with HSW, an exponential weighting w_t ∝ λ^{t-1}, a recursion-depth curriculum, amortized candidate generation, and early stopping. The paper reports strong accuracies (e.g., 88.5% Sudoku, 93.1% Maze, 82.1% ARC-1, 47.8% ARC-AGI) and claims consistent gains over prior recursive models and much larger zero-shot LLMs.

Significance. The paper's strengths include a clean unified formulation of diverse symbolic tasks as serialized constraint satisfaction, an explicit constraint-preserving projection, parameter-efficient LoRA adaptation, and a fair controlled comparison against HRM/TRM/LoopFormer trained from scratch. The Figure 5 ablation provides useful evidence that recursion-aware training, rather than mere test-time self-revision, drives the improvement. However, the headline claim of outperforming 'substantially larger LLMs' rests on unmatched zero-shot leaderboard comparisons, the marginal contribution of recursion over a non-recursive LoRA fine-tuned baseline is not reported in the main tables, and the '50% gradient-variance reduction' attributed to HSW is never measured. These issues are load-bearing for the paper's central claims and need to be resolved.

major comments (4)
  1. [Abstract; §4, Tables 1–3] The headline claim that R-Qwen 'outperforms substantially larger LLMs while using a comparable number of trainable parameters' is not supported by matched comparisons. Table 1's LLM numbers are zero-shot leaderboard scores, whereas R-Qwen is trained or adapted on each benchmark's training data; the text acknowledges this distinction only in a footnote/caption. The 'comparable trainable parameters' framing is also misleading because the frozen 27B backbone contributes substantial pretraining priors, whereas the from-scratch recursive baselines have 7–278M total parameters. A non-recursive LoRA fine-tuned model (one refinement step) should be reported alongside Tables 1–3, and matched fine-tuning comparisons to open-weight LLMs are needed to support the headline outperformance claim.
  2. [Abstract; §3.3, Eqs. (7)–(8)] The abstract and §3.3 state that HSW 'reduces gradient variance by at least 50%' and improves the signal-to-noise ratio of stochastic gradients. No measurement of gradient variance or SNR appears anywhere in the paper; the only supporting evidence is training-time speedup in Table 4, which is confounded with the curriculum and early-stopping mechanisms. Moreover, λ in Eq. (8) is a free parameter and no ablation over λ is reported. The authors should either measure the claimed variance reduction directly or remove the quantitative claim.
  3. [Algorithm 1; §3.3] There is a mismatch between the aggregated HSW objective in Eq. (8) and the implemented training loop. Algorithm 1 lines 16–26 perform one optimizer step per refinement step, scaling the loss by w_t. Per-step SGD updates of the form θ_t = θ_{t-1} - η w_t ∇ℓ_t are not equivalent to a single aggregated update on Eq. (8); they implement a per-step learning-rate modulation. The statement that accumulating weighted losses and applying a single update 'recovers Eq. (8) exactly' is true only for the aggregation variant, not for the per-step loop actually used. The paper should specify which objective is optimized and reconcile the HSW justification with the implementation.
  4. [Figure 5; §4 Ablations] The ablation text says 'simply applying recursive refinement to a pretrained VLM at inference time (Experiment 1) does not improve performance,' but the Figure 5 caption says the curves indicate 'the number of recursion steps used during training.' It is unclear whether Experiment 1 is a zero-shot frozen model or a LoRA model trained with one recursion step. More importantly, the main tables omit the accuracy of the fine-tuned backbone without recursion. Without that number, the reported gains (e.g., 88.5% Sudoku, 82.1% ARC-1) cannot be decomposed into pretraining/fine-tuning improvements versus the contribution of recursive refinement. Please report the one-step (or zero-step) fine-tuned baseline on each benchmark and clarify the status of Experiment 1.
minor comments (6)
  1. [Abstract; Table 2] The abstract's '27.6% improvement over the baseline' is undefined. Table 2 shows different improvements depending on the baseline chosen (5.6 vs o4-mini, 25.8 vs VLSR, 34.4 vs LoopFormer). Specify which baseline is meant.
  2. [Table 1 caption] The caption states 'LLM-based results are from the ARC-AGI leaderboard,' which is implausible for Sudoku and Maze. Clarify the source of each column and whether the metric is position-level accuracy or exact-match accuracy.
  3. [§3.2; §Methods] There are missing/incorrect cross-references: 'Supplementary Section .' and 'LoRa Section.' Also, the value of the HSW decay parameter λ used in the experiments is never stated. A λ ablation would be useful even as a small table.
  4. [Figures 3–5] The curves in Figures 3–5 are single runs without error bars or seed information. Given that the cross-word gains are sometimes only 1–2%, reporting variance would materially increase confidence in the claims.
  5. [Table 4] In the caption of Table 4, the 'standard model' baseline is not defined. If it means uniform deep supervision, state this explicitly and describe how HSW, Halting Act, and the standard model are otherwise matched in compute.
  6. [Abstract] The GitHub URL appears in the abstract, but the text also says 'Code and models will be released after acceptance.' Clarify the actual availability of code and model weights.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: benchmark gains are empirical, not derived from the HSW weighting or LoRA fit; the one self-citation is non-load-bearing.

full rationale

I find no circular step in the derivation chain. The training objective is constructed from Eq. (2) (generation-projection recurrence), Eq. (4) (step-wise cross-entropy), and Eq. (8) (HSW objective). Eq. (8) contains a free hyperparameter λ and reduces to uniform deep supervision at λ=1; no parameter is fitted to a benchmark and then renamed as a prediction. The headline accuracy numbers in Tables 1-3 are measured outcomes of training and inference, not consequences of any equation in the paper. The constraint projection Π_x in Eq. (3) does guarantee by construction that clamped cells are preserved and invalid outputs fall back to the previous candidate, but that does not force the whole-solution exact-match results, and the fixed cells are not presented as model predictions. The only self-citation is Venkatraman et al. 2025 (Lajoie is a co-author), used only as a related-work contrast: "Unlike inference-time recursive schemes that recurse over the prompt (Zhang, Kraska, and Khattab 2025; Venkatraman et al. 2025) or looped models with learned elastic depth..." It is not load-bearing, and no uniqueness or optimality claim rests on it. The abstract's statement that "HSW reduces gradient variance by at least 50%" is not measured in this paper and is imported from CGAR (Qasim and Zhang 2025); this is a missing-support / correctness-risk issue, not circularity, because the Table 4 speedup/accuracy comparison is an empirical measurement and the benchmark results do not depend on the variance-reduction premise. The paper itself flags the main HSW attribution gap: "the comparison in Figure 4 does not isolate its effect. Determining whether this inference-time robustness is specifically attributable to HSW requires a controlled ablation against uniform deep supervision, which is a subject of future work." That limitation weakens evidential strength but does not make the derivation circular. The score is 2 rather than 0 only because one minor, non-load-bearing self-citation is present; the central derivation is self-contained and empirically evaluated.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central results rest on a handful of free hyperparameters (notably lambda and the chosen recursion depth) and on the assumption that deep supervision plus truncated recurrence trains a stable refinement operator. No new physical or mathematical entities are introduced.

free parameters (4)
  • HSW decay rate lambda = not reported
    Equation (7) defines the exponential weighting; the value of lambda is never given in the main text or appendix, yet it controls the HSW emphasis. Likely tuned but unreported.
  • Recursion depth for main experiments = 8 steps
    Selected from the ablation in Figure 3; the paper states 'we use 8 recursive steps in our main experiments as a good trade-off'.
  • Supervision-depth curriculum schedule = (1:4, 4:8, 7:10)
    Hand-specified in Table 6 of the appendix; no rule is given for choosing these epochs.
  • Candidate generation interval m = not reported
    Algorithm 1 uses an interval m for amortised decoding; the value is not specified in the text.
assumptions (4)
  • domain assumption Teacher-forcing the full target y* at every step teaches a refinement operator that is stable under greedy decoding at inference
    Eq. (4) supervises each step with the complete target; the paper assumes this generalizes to the candidate distribution seen when iterating Eq. (2) at test time.
  • domain assumption Truncated recurrence with detached candidate states preserves learning signal
    Candidates are detached and fed as fixed inputs, following TRM; the approximation is stated but not analyzed.
  • standard math The constraint projection Pi_x guarantees a valid state and never destroys the problem structure
    Eq. (3) clamps fixed positions, validates the alphabet, and truncates or pads to length L; this is deterministic.
  • domain assumption The variance-reduction benefit of HSW shown for tiny recursive models transfers to a 27B autoregressive LLM
    The paper cites CGAR (Qasim and Zhang 2025) for 'variance reduction' and 'signal-to-noise' but does not measure either for R-Qwen.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Recursive Vision Language Models for General Symbolic Reasoning." pith.science (2026). https://pith.science/paper/HTTRPCOZ

@misc{pith2026260801534,
  author       = {Pith},
  title        = {Pith review of: Recursive Vision Language Models for General Symbolic Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HTTRPCOZ}},
  note         = {Machine review of arXiv:2608.01534}
}
read the original abstract

Hard symbolic-reasoning tasks such as Sudoku, maze pathfinding, and ARC remain challenging for LLMs due to their fixed-depth autoregressive reasoning, which limits systematic search, refinement, and backtracking. While recursive models such as Hierarchical Reasoning Model (HRM) and Tiny Recursive Model (TRM) address this limitation through iterative latent-state refinement, they are typically task-specific and do not leverage pretrained language priors. We propose R-Qwen, a recursive reasoning framework built upon a pretrained Qwen backbone. R-Qwen repeatedly refines a candidate solution through programmatic self-recursion and deep supervision, combining the structured iterative computation of recursive models with the linguistic and reasoning priors of pretrained LLMs. We further adapt Hierarchical Supervision Weighting (HSW) to autoregressive models by exponentially weighting losses across recursive steps. HSW reduces gradient variance by at least 50\%, improves the signal-to-noise ratio of stochastic gradients, and accelerates convergence. Across eight challenging benchmarks, R-Qwen consistently outperforms prior recursive reasoning models and substantially larger LLMs while using a comparable number of trainable parameters. Notably, on ARC-AGI dataset, our model achieves a 27.6\% improvement over the baseline, highlighting the effectiveness of recursive refinement for general symbolic reasoning. These results suggest that recursive reasoning mechanisms and pretrained language model priors are complementary approaches for improving symbolic puzzle-solving. Code and models will be released after acceptance.

Figures

Figures reproduced from arXiv: 2608.01534 by the authors.

Figure 1
Figure 1. The proposed recursive reasoning framework with [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The proposed task-general recursive candidate refinement framework. Each structured reasoning task is serialized into [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Accuracy comparison across recursion steps on three benchmarks. Our method achieves substantially higher accuracy [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Exact accuracy across recursion steps. The right [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Exact accuracy across recursion steps. The right [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Visualization of a Sudoku puzzle and its solving steps using our R-Qwen model. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Visualization of a Maze puzzle and its solving steps using our R-Qwen model. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Visualization of a ARC puzzle and its solving steps using our R-Qwen model. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 16 canonical work pages

  1. [4]

    Chollet, F

    Bring reason to vision: Understanding perception and reasoning through model merging.arXiv preprint arXiv:2505.05464. Chollet, F

  2. [6]

    Boosting Inference with Guided Reasoning: Stochastic Exploration for Recursive Models

    Boosting Inference with Guided Rea- soning: Stochastic Exploration for Recursive Models.arXiv preprint arXiv:2605.25230. Dehghani, M.; Gouws, S.; Vinyals, O.; Uszkoreit, J.; and Kaiser, L

  3. [7]

    InInternational Conference on Learning Representations

    Universal Transformers. InInternational Conference on Learning Representations. Dettmers,T.;Pagnoni,A.;Holtzman,A.;andZettlemoyer,L. 2023.Qlora:Efficientfinetuningofquantizedllms.Advances inneuralinformationprocessingsystems,36:10088–10115. Dong,Y.;Liu,Z.;Sun,H.-L.;Yang,J.;Hu,W.;Rao,Y.;and Liu,Z.2025. Insight-v:Exploringlong-chainvisualreason- ing with mu...

  4. [8]

    Bridging the Gap Between Latent and Explicit Reasoning with Looped Transformers

    Bridging the Gap Between Latent and Explicit Reasoning with Looped Trans- formers.arXiv preprint arXiv:2606.31779. Fein-Ashley, J.; and Rashidinejad, P

  5. [9]

    Freinschlag,R.;Bertram,T.;Kobler,E.;Mayr,A.;andKlam- bauer, G

    Solve the Loop: Attractor Models for Language and Reasoning.arXiv preprint arXiv:2605.12466. Freinschlag,R.;Bertram,T.;Kobler,E.;Mayr,A.;andKlam- bauer, G

  6. [10]

    InProceedings of the 43rd International Confer- ence on Machine Learning, Proceedings of Machine Learn- ing Research

    Symbol-Equivariant Recurrent Reasoning Models. InProceedings of the 43rd International Confer- ence on Machine Learning, Proceedings of Machine Learn- ing Research. PMLR. ArXiv:2603.02193. Fu, R.; and Luo, G

  7. [11]

    InProceedingsofthe64thAnnual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 7073–7087

    Selar: Selective latent reasoning inlargelanguagemodels. InProceedingsofthe64thAnnual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 7073–7087. Geiping, J.; McLeish, S.; Jain, N.; Kirchenbauer, J.; Singh, S.;Bartoldson,B.;Kailkhura,B.;Bhatele,A.;andGoldstein, T.2025. ScalingupTest-TimeComputewithLatentReason- ing: A Recu...

  8. [12]

    Hammoud, H

    Form Follows Function: Recursive Stem Model.arXiv preprint arXiv:2603.15641. Hammoud, H. A. A. K.; Alhamoud, K.; Hammoud, A.; Bou-Zeid, E.; Ghassemi, M.; and Ghanem, B

Show all 39 references
  1. [13]

    arXiv preprint arXiv:2508.08940

    Train long,thinkshort:Curriculumlearningforefficientreasoning. arXiv preprint arXiv:2508.08940. Hao,S.;Sukhbaatar,S.;Su,D.;Li,X.;Hu,Z.;Weston,J.;and Tian,Y.2025. Traininglargelanguagemodelstoreasonina continuous latent space.arXiv preprint arXiv:2412.06769. He, J.; Lin, H.; Wa...

  2. [14]

    Hodel, M

    Flow Reasoning Models: Scaling Rea- soning Through Iterative Self-Refinement.arXiv preprint arXiv:2606.29150. Hodel, M

  3. [16]

    InInternational conference on learn- ing representations, volume 2024, 32808–32824

    Large language models cannot self- correct reasoning yet. InInternational conference on learn- ing representations, volume 2024, 32808–32824. Jeddi, A.; Ciccone, M.; and Taati, B

  4. [17]

    InInternational Conference on Learning Representations (ICLR)

    LoopFormer: Elastic-Depth Looped Transformers for Latent Reasoning via Shortcut Modulation. InInternational Conference on Learning Representations (ICLR). ArXiv:2602.11451. Jolicoeur-Martineau, A

  5. [18]

    ArXiv:2510.04871, arXiv:2510.04871

    Less is More: Recur- sive Reasoning with Tiny Networks. ArXiv:2510.04871, arXiv:2510.04871. Kohli,H.;Parthasarathy,S.;Sun,H.;andYao,Y.2026.Loop, think, & generalize: Implicit reasoning in recurrent-depth transformers.arXiv preprint arXiv:2604.07822. Komisarczyk, M.; Mathur, S....

  6. [19]

    Leng, J.; Huang, C.; Huang, L.; Lin, B

    Beyond a*: Better planning with transformers via search dynamics bootstrapping.arXiv preprint arXiv:2402.14083. Leng, J.; Huang, C.; Huang, L.; Lin, B. Y.; Cohen, W. W.; Wang,H.;andHuang,J.2025. Crosswordbench:Evaluating thereasoningcapabilitiesofllmsandlvlmswithcontrollable p...

  7. [20]

    InInternational Conference on Learning Representations, volume 2025, 20981–21024

    Combininginductionandtransductionforabstractreasoning. InInternational Conference on Learning Representations, volume 2025, 20981–21024. Madaan, A.; Tandon, N.; Gupta, P.; Hallinan, S.; Gao, L.; Wiegreffe, S.; Alon, U.; Dziri, N.; Prabhumoye, S.; Yang, Y.; et al

  8. [22]

    Palm, R.; Paquet, U.; and Winther, O

    MaxSAT- Based Feedback for Guiding Vision-Language Models in Sudoku.arXiv preprint arXiv:2607.12711. Palm, R.; Paquet, U.; and Winther, O

  9. [24]

    Rauba, P.; Fanconi, C.; and van der Schaar, M

    Accelerating Training Speed of Tiny Recursive Models with Curriculum Guided Adaptive Recursion.arXiv preprint arXiv:2511.08653. Rauba, P.; Fanconi, C.; and van der Schaar, M

  10. [25]

    Tiny Autoregressive Recursive Models.arXiv preprint arXiv:2603.08082. Ren, H

  11. [26]

    In Findings of the Association for Computational Linguistics: EMNLP 2025, 12534–12545

    LSRL: Process-Supervised GRPO on La- tent Recurrent States Improves Mathematical Reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2025, 12534–12545. Ren,Z.;andLiu,Z.2026. AreYourReasoningModelsRea- soningorGuessing?AMechanisticAnalysisofHierarchic...

  12. [27]

    Sharma, R.; and Vu, T

    Probabilistic Tiny Recursive Model.arXiv preprint arXiv:2605.19943. Sharma, R.; and Vu, T

  13. [28]

    Tang, G.; Jiang, S.; Chang, H.; Chen, N.; Li, Y.; Fan, H.; Li, J.; Liu, M.; and Qin, B

    Dense Supervision Is Not Enough:TheReadoutBlindSpotinLoopedLanguageMod- els.arXiv preprint arXiv:2606.24898. Tang, G.; Jiang, S.; Chang, H.; Chen, N.; Li, Y.; Fan, H.; Li, J.; Liu, M.; and Qin, B

  14. [29]

    Venkatraman, S.; Jain, V.; Mittal, S.; Shah, V.; Obando- Ceron, J.; Bengio, Y.; Bartoldson, B

    Looprpt: Reinforcement pre-training for looped language models.arXiv preprint arXiv:2603.19714. Venkatraman, S.; Jain, V.; Mittal, S.; Shah, V.; Obando- Ceron, J.; Bengio, Y.; Bartoldson, B. R.; Kailkhura, B.; La- joie, G.; Berseth, G.; et al

  15. [30]

    Wang, G.; Li, J.; Sun, Y.; Chen, X.; Liu, C.; Wu, Y.; Lu, M.; Song, S.; and Abbasi Yadkori, Y

    Recursive self-aggregation unlocks deep thinking in large language models.arXiv preprint arXiv:2509.26626. Wang, G.; Li, J.; Sun, Y.; Chen, X.; Liu, C.; Wu, Y.; Lu, M.; Song, S.; and Abbasi Yadkori, Y

  16. [31]

    SpiralThinker:LatentReasoning through an Iterative Process with Text-Latent Interleaving

    Piao,S.;andPark,S.2025. SpiralThinker:LatentReasoning through an Iterative Process with Text-Latent Interleaving. arXiv preprint arXiv:2511.08983. Qasim, K. U.; and Zhang, J

  17. [32]

    ArXiv:2506.21734, arXiv:2506.21734

    Hierarchical Reasoning Model. ArXiv:2506.21734, arXiv:2506.21734. Wang, W.; and Reid, F

  18. [33]

    Wei, A.; Wu, Y.; Wan, Y.; Suresh, T.; Tan, H.; Zhou, Z.; Koyejo,S.;Wang,K.;andAiken,A.2025

    Tiny recursive reasoning with mamba-2attentionhybrid.arXivpreprintarXiv:2602.12078. Wei, A.; Wu, Y.; Wan, Y.; Suresh, T.; Tan, H.; Zhou, Z.; Koyejo,S.;Wang,K.;andAiken,A.2025. Satbench:Bench- markingllms’logicalreasoningviaautomatedpuzzlegener- ation from sat formulas. InProce...

  19. [34]

    Yang, C.; Srebro, N.; and Li, Z

    Looping back to move forward: Recursive transformers for efficient and flexible large multi- modal models.arXiv preprint arXiv:2602.09080. Yang, C.; Srebro, N.; and Li, Z

  20. [35]

    Zhang, A

    Recursive models for long-horizon reasoning.arXiv preprint arXiv:2603.02112. Zhang, A. L.; Kraska, T.; and Khattab, O

  21. [36]

    Zhang, B.; Zang, Y.; Dong, X.; Cao, Y.; Duan, H.; Lin, D.; andWang,J.2026.ThinkVisually,ReasonTextually:Vision- LanguageSynergyinAbstractReasoning

    Recursive language models.arXiv preprint arXiv:2512.24601. Zhang, B.; Zang, Y.; Dong, X.; Cao, Y.; Duan, H.; Lin, D.; andWang,J.2026.ThinkVisually,ReasonTextually:Vision- LanguageSynergyinAbstractReasoning. InProceedingsof the IEEE/CVF Conference on Computer Vision and Pattern...

  22. [37]

    Energy-guided Recursive Model.arXiv preprint arXiv:2607.10128. Supplementary Material Extended Related Work Recursivereasoningmodels.HRMusestworecurrentmod- ules operating at different timescales, while TRM simplifies thisdesigntoasinglenetworkthatrepeatedlyrefinesalatent pred...

  23. [38]

    R-Qwen instead uses a pretrained autore- gressive backbone and represents each recurrent state as an explicit serialized candidate

    also require a special- ized position–symbol representation whose cost grows with the alphabet size. R-Qwen instead uses a pretrained autore- gressive backbone and represents each recurrent state as an explicit serialized candidate. TheFixed-PointReasoningModelstabilizeslooped...

  24. [39]

    similarly operate over latent trajectories or latent tokens and modify the network architecture. RecursiveVLM adapts parameter-shared recur- sionspecificallytolargemultimodalmodelsthrougharecur- sive connector that aligns intermediate vision and language features and a monoton...

  25. [40]

    Using 16 particles and 48 outer recur- sionsteps,GuidedMAPimprovesSudoku-Extremeaccuracy from 85.9% to 98.0%

    apply guided stochas- tic exploration to a frozen TRM model without retrain- ing its backbone. Using 16 particles and 48 outer recur- sionsteps,GuidedMAPimprovesSudoku-Extremeaccuracy from 85.9% to 98.0%. However, the learned guide does not provideacorrespondingimprovementonMa...

  26. [2019]

    ArXiv:1911.01547, arXiv:1911.01547

    On the Measure of Intelligence. ArXiv:1911.01547, arXiv:1911.01547. Chollet, F.; and ARC Prize Foundation

  27. [2022]

    Behzadian,B.;Daley,B.;Subbaraj,G.;andNassif,H.2026

    End-to- endalgorithmsynthesiswithrecurrentnetworks:Extrapola- tion without overthinking.Advances in Neural Information Processing Systems, 35: 20232–20242. Behzadian,B.;Daley,B.;Subbaraj,G.;andNassif,H.2026. Unrolled Policy Iteration for Tiny Recursive Models.ICLR Workshop. Ch...

  28. [2023]

    Movahedi, S.; Milovanović, V.; Feigin, S

    Self-refine: Iterative refinement with self- feedback.Advances in neural information processing sys- tems, 36: 46534–46594. Movahedi, S.; Milovanović, V.; Feigin, S. L.; Theus, A.; Hofmann,T.;Boeva,V.;Rusch,T.K.;andOrvieto,A.2026. Fixed-Point Reasoners: Stable and Adaptive Dee...

  29. [2024]

    Hu,E.J.;Shen,Y.;Wallis,P.;Allen-Zhu,Z.;Li,Y.;Wang,S.; Wang, L.; Chen, W.; et al

    Addressing the abstraction and reasoning corpus via procedural example generation.arXiv preprint arXiv:2404.07353. Hu,E.J.;Shen,Y.;Wallis,P.;Allen-Zhu,Z.;Li,Y.;Wang,S.; Wang, L.; Chen, W.; et al

  30. [2025]

    InPro- ceedings of the 42nd International Conference on Machine Learning, volume 267 ofProceedings of Machine Learning Research, 942–963

    The Surprising Effective- ness of Test-Time Training for Few-Shot Learning. InPro- ceedings of the 42nd International Conference on Machine Learning, volume 267 ofProceedings of Machine Learning Research, 942–963. PMLR. Asadulaev,A.;Banerjee,R.;Karray,F.;andTakáč,M.2026. Laten...

  31. [2026]

    Bansal, A.; Schwarzschild, A.; Borgnia, E.; Emam, Z.; Huang, F.; Goldblum, M.; and Goldstein, T

    Generative recursive reasoning.arXiv preprint arXiv:2605.19376. Bansal, A.; Schwarzschild, A.; Borgnia, E.; Emam, Z.; Huang, F.; Goldblum, M.; and Goldstein, T

Pith tools

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