Disentangling Perception and Reasoning in Multimodal LLMs via Reward Design

Eftekhar Hossain; Nikhil Singh; Omar Sharif; Patrick Ng

REVIEW 4 major objections 6 minor 1 cited by

Disentangling Perception and Reasoning in Multimodal LLMs via Reward Design

T0 review · 4 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read This paper argues that on algorithmic visual puzzles, multimodal language models fail mainly because they cannot read the image reliably, not because they cannot reason about it, and that reward-driven reinforcement learning can induce long

desk verdict The perception-bottleneck result is plausible but the text-substitution test is confounded by algorithm-friendly encodings; the reward ablation is systematic but under-powered — worth a serious referee, with major revision. read the letter →

arxiv 2601.00215 v2 pith:XJHDCY3Y submitted 2026-01-01 cs.CV cs.CL

Omar Sharif , Eftekhar Hossain , Nikhil Singh , Patrick Ng This is my paper · ORCID

classification cs.CVcs.CL

keywords visualperceptionbottleneckmultimodallargelanguagemodelsreinforcementlearningrewarddesignreasoningalgorithmicpuzzlesGRPOAlgoPuzzleVQA

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 is trying to establish that the primary bottleneck for multimodal LLMs on visual puzzles is perception, not algorithmic reasoning. It tests this by replacing images with hand-built text descriptions: across nine puzzle categories, Claude 3.5 improves by 26.7 percentage points and Claude 3.7 by 23.6 points, with categories like Clock and N-Queens jumping from near-failure to near-perfect once the visual step is removed. The paper then asks whether reinforcement learning can improve the perception side without costly chain-of-thought supervision, and shows that a mixture reward—paying separately for image description, thinking, rethinking, and exact answer correctness—trained Qwen-2.5-VL-7B to reach 15.56% average accuracy versus a 10% baseline, while also producing long, structured, self-checked reasoning. The gains are real but uneven: no single reward function wins across all categories, and some rewards trade out-of-domain transfer for in-domain accuracy. A sympathetic reader would take the central contribution as causal evidence that perception, not reasoning, is the binding constraint, with reward design as a partial remedy.

What carries the argument

The central mechanism is a controlled image-to-text substitution: a rule-based mapping that replaces each puzzle image with a minimal text representation (clock times as '2:43', checker boards as strings, mazes as 0/1 grids, queen positions as coordinate lists) so the algorithmic content is preserved while the perceptual step is removed. The second mechanism is a family of reward functions for GRPO training. All rewards route outputs through structured tags, and the mixture and continuous variants score the amount of unique content inside <image description>, <think>, and <rethink> tags using tanh-smoothed counts with repetition penalties, weighted heavily toward answer accuracy (alpha_i=0.0

What would settle it

Take a category with a large text-only gain, such as N-Queens or Clock, and feed the model two text forms of the same puzzle: a human-friendly coordinate list and an equally complete but harder-to-parse form (for example, a rotated or color-scrambled grid that still demands visual-like parsing). If accuracy stays high in both, the perception-bottleneck reading is supported; if accuracy drops sharply in the harder-to-read form, then the original gain came at least partly from the encoding being algorithmically convenient rather than from removing a perceptual step.

Watch

Extended reading notes

Core claim

The paper establishes two results. First, using nine categories of the AlgoPuzzleVQA visual-puzzle benchmark, converting images into character- or coordinate-based text descriptions raises average open-ended accuracy from 25.9% to 52.6% for Claude 3.5 and from 42.4% to 66.0% for Claude 3.7, with near-perfect scores in N-Queens and large gains in Clock. Because every other component of the prompt is unchanged, the paper interprets this as evidence that visual perception is the primary bottleneck. Second, training Qwen-2.5-VL-7B with GRPO under six reward designs shows that a mixture reward—which separately rewards image description, thinking, rethinking, and final-answer correctness—induces l

Load-bearing premise

The claim that perception is the binding constraint rests on the assumption that the hand-crafted text descriptions are exact perceptual equivalents—that they remove only the image-reading step and leave the algorithmic difficulty of the puzzle unchanged.

Editorial extensions

If this is right

Improving visual perception is the highest-leverage direction: fixing image reading would unlock larger accuracy gains than improving algorithmic reasoning.
Reward signals can elicit long, visually grounded reasoning in open-weights multimodal LLMs without dense chain-of-thought supervision, so RL-based visual reasoning can scale without costly annotations.
Reward design has no universal winner: no single reward function is best for all categories, and accuracy-focused rewards trade out-of-domain generalization for in-domain performance.
Open-ended evaluation is preferable to multiple-choice in visual puzzle benchmarks, since multiple-choice lets models guess from options and can mask reasoning failures.
Some puzzle classes, such as Clock and Tower of Hanoi, remain hard even after RL training, indicating that abstract temporal and recursive demands are not addressed by current reward designs.

Reading between the lines

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

The bottleneck claim likely overstates pure perception: the hand-built text encodings are algorithm-friendly (coordinate lists, grid arrays) that also lower reasoning difficulty, so part of the 20-plus-point gain may be representation convenience rather than bypassed perception. A stricter test would use text encodings matched in difficulty, such as rotated or obfuscated grids.
If perception is the binding constraint, then perception-aware reward signals—for example, rewarding accurate extraction of object positions and attributes using a separate judge—should outperform length-based rewards. The paper itself suggests an LLM-based judge for intermediate reasoning as a future direction.
The mixture reward's long self-checked reasoning may partly be an artifact of rewarding unique content volume: a reward that pays per unique sentence encourages verbosity without necessarily improving grounding, which would explain why accuracy gains are modest relative to reasoning length. A controlled comparison with a fixed-budget reasoning reward could test this.
The text-substitution protocol could serve as a cheap diagnostic for any multimodal benchmark: measure the text-only ceiling first, and if it is much higher than the image-based score, the bottleneck is perception, and training should target perception before reasoning.

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.

Desk Editor's Note

read the letter

Two things to know. First, the bottleneck experiment is the most interesting part, but the text-only condition isn't a clean control. Second, the RL reward ablation is systematic but the empirical evidence is too thin to support the strong claims.

The paper's core contribution is the text-substitution test on AlgoPuzzleVQA: swapping images for hand-written textual descriptions lifts Claude 3.5 from 25.9% to 52.6% and Claude 3.7 from 42.4% to 66.0%. That's a big effect, and the qualitative errors in Table 8 show real perception failures propagating into reasoning. The six-reward ablation is also a reasonable first map of the reward-design space, with clear pseudo-code and honest reporting that no single reward wins across categories.

The soft spots are in proportion. The bottleneck claim is confounded by the representation itself. Table 7's encodings do more than remove perception: for Clock they hand you the exact time (2:43), for N-Queens they give coordinate lists that also skip the spatial indexing step. The categories are selected for text-convertibility, so the comparison is biased toward tasks where the text form is algorithm-friendly. The 20+ point gains could partly come from a simpler reasoning problem. The stress-test note is right that this is load-bearing: the Section 5.1 conclusion overstates what the data show.

The RL half is promising as a design study but weak as evidence. Single runs, no error bars, per-category samples that look small, hand-tuned weights, and no released code or data. The mixture reward's "long reasoning" is partly circular — the pseudo-code in Table 10 shows the reward is a function of unique sentences in the tags, so the model is doing what it was rewarded to do. That's not a flaw in the reward design per se, but "unlocks" is the wrong verb. The claim that long reasoning improves perception is also not directly measured; they report reasoning length and final accuracy, not perception accuracy.

If I were the editor I'd send it out: the bottleneck question matters, and the confound is addressable with better controls — for example, text forms that preserve the same perceptual difficulty, or a noisier text condition. The authors are honest about the 15.56% ceiling and the unevenness. This needs major revision, but it deserves referee time.

Referee Report

4 major / 6 minor

Summary. The paper studies visual puzzle solving in multimodal LLMs. It first claims that visual perception, not algorithmic reasoning, is the binding constraint: replacing images with hand-crafted textual representations raises Claude 3.5 accuracy by 26.7 points and Claude 3.7 by 23.6 points (Table 3). It then proposes six reward functions for GRPO training of Qwen-2.5-VL-7B, reporting that a 'mixture' reward achieves 15.56% average accuracy versus roughly 10% for the base model, while producing long, structured, self-reflective reasoning. The paper also reports in-domain/out-of-domain generalization results and a qualitative error analysis. The bottleneck evidence is quantitatively striking but confounded by the choice of text encodings; the RL results are useful but limited by reward design circularity, single runs, and small per-category samples.

Significance. If the bottleneck claim were cleanly established, it would be an important result for multimodal reward design and for diagnosing where MLLMs fail on visual puzzles. The paper contributes a systematic taxonomy of six reward functions, detailed pseudo-code (Table 10), open-ended evaluation, and qualitative examples of perception-driven errors. These are useful resources for the community. However, the central causal claim—that perception, not reasoning, is the binding constraint—is not yet established because the text-only condition also simplifies the reasoning representation. The paper's secondary claims about reward-induced long reasoning are partially circular and lack statistical validation. The significance is therefore conditional on a re-scoped or better-controlled bottleneck experiment.

major comments (4)

[§5.1, Tables 3 and 7] The text-only condition is not a perception-free control. In the Clock category, the image is reduced to 'Current Time: 2:43', i.e., the exact perceptual read-out, leaving only arithmetic. In N-Queens, queen positions are given as coordinate lists [0,6], [1,3], ... , removing both queen detection and coordinate assignment. Maze and Move Box are rendered as 0/1 or character arrays, which also eliminate visual grouping, counting, and spatial tracing. Since §3 selects categories 'based on their suitability for conversion into text-only representations,' the comparison conflates removing perception with reducing representational/reasoning difficulty. The +26.7/+23.6 pp gains therefore do not establish that perception is the primary bottleneck. A control that preserves visual information format (e.g., visual grounding annotations or spatial-coordinate supervision) or a re-scoped claim is need
[§5.3, Table 4] The statement 'OOD performance with RL consistently exceeds the no-RL baseline' is contradicted by Table 4. In the Clock-3K setting, Vanilla OOD is 4.25, Only-Accuracy 9.37, Continuous 10.12, and Visual-Fusion 9.12, all below the 10.75 baseline. In the Diverse-8k setting, Vanilla OOD is 8.0, below the 9.4 baseline. The abstract's claim of 'consistent gains across both in-domain and out-of-domain settings' is therefore unsupported. This is not a presentation issue: the generalization claim is one of the stated contributions.
[§2.2, Table 10] The mixture, continuous, and no-accuracy rewards include SideReward, which scores tanh(unique sentences) - tanh(duplicates) for <image description>, <think>, and <rethink> tags. Observing long, structured reasoning after RL is therefore partly by construction: the reward is directly proportional to the number of unique sentences in these tags. The paper presents this as 'unlock[ing]' emergent reasoning (§5.2), but the behavior is the optimization target. A length-controlled baseline or a content-quality metric independent of the rewarded signal would strengthen the claim. In addition, all RL results are single runs without error bars; with per-category test sizes near 100, class-level differences such as 1 vs 37 for Clock have wide confidence intervals, and the reported average gains may not be stable.
[§4.2, Table 9 and Appendix A] The reward-function comparison is partly confounded by prompt differences. Vanilla and Visual-Fusion use separate prompts, while Only-Accuracy, Mixture, Continuous, and No-Accuracy share a different prompt. Thus some differences may reflect prompt format rather than reward computation. Additionally, the reward weights αi, αt, αr, αv, αa were selected by trial-and-error, and no sensitivity analysis is reported. This limits conclusions about which reward component is responsible for the observed gains and about the generality of the reward design recommendations.

minor comments (6)

[Appendix A] Typo: 'there for excluded' should be 'therefore excluded'. In Table 10, 'Retrun' should be 'Return'; in Table 7, 'capcities' should be 'capacities'.
[Table 2] The table formatting appears broken: the Diverse-8k baseline column is missing or misaligned, and row lengths are inconsistent. Please fix the column structure so that per-category baseline values for both splits are clear.
[Table 9] The text says Only-Accuracy, Mixture, Continuous, and No-Accuracy share the same prompt, yet the reward contents differ. Please clarify whether all tags are always generated by the model and only the reward computation changes.
[§2.1, Table 1] The No-Accuracy row says 'full credit for generating an <answer> tag' in the text, but Table 1 lists the reward content as 'answer-tag'. Nomenclature should be consistent.
[References] Duplicate references for GRIT (Fan et al. 2025a/2025b) and DeepEyes (Zheng et al. 2025b/2025c) should be consolidated.
[General] The arXiv metadata title 'Disentangling Perception and Reasoning...' differs from the in-manuscript title 'From Sight to Insight...'. Please align the metadata.

Circularity Check

1 steps flagged · score 4.0 of 10

Long-reasoning 'unlock' is a direct consequence of the reward definition; the perception-bottleneck experiment is not circular.

self definitional [Table 1; Table 10 (SideReward); Figure 15 caption; Section 5.2]
"Each non-answer tag’s score is based on the amount of unique content generated within that tag, with counts passed through a tanh() smoothing function... SideReward: ... COUNT unique_sentences ... smoothed_score <- tanh(unique_sentences) ... / The designed reward encourages long, detailed chains of thought, yielding clear, stepwise, and self-checked explanations."

The mixture/continuous/no-accuracy rewards pay explicitly for the number of unique sentences inside <image description>, <think>, and <rethink> tags (tanh(unique_sentences)). Observing that the trained model produces long, structured, self-checked reasoning is therefore the optimization target being realized, not an emergent property discovered by the experiments. The paper's framing — 'Mixture reward unlocks long reasoning capability' — reduces by construction to 'the reward was defined to reward unique tagged content.' The accuracy gain of 5.56 points is separate and not circular, but the long-reasoning 'unlock' claim is the reward function restated.

full rationale

The central bottleneck claim (Section 5.1) is not circular: it is an external comparison between multimodal and text-only accuracy on AlgoPuzzleVQA, with hand-crafted textual representations. One can question whether those representations also reduce algorithmic difficulty (e.g., Clock becomes 'Current Time: 2:43', N-Queens becomes coordinate lists), but that is a validity/confound concern, not an equation-level reduction, and the result is externally falsifiable rather than resting on self-citation. The reward-design portion, however, contains one self-definitional step: the mixture and related rewards score unique sentence counts within the reasoning tags, so the reported 'emergence' of long visual reasoning is directly incentivized by construction. This is partial circularity in one of the paper's secondary claims, while the main perception-bottleneck result and the measured accuracy gains remain independent. No load-bearing self-citation chain or imported uniqueness theorem is present. Overall score 4 reflects that partial circularity rather than a fully forced derivation.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

Central claims rest on three categories of external input: (1) the AlgoPuzzleVQA benchmark as a valid set of labels and questions; (2) a set of hand-designed text encodings that the paper assumes are lossless; and (3) a GRPO/EasyR1 implementation plus hand-tuned reward weights and continuous-reward thresholds. The bottleneck claim is most sensitive to (2), since category selection was biased toward text-convertible puzzles. The RL claims are sensitive to (3), because no sensitivity analysis or multiple seeds is reported, and the reward for unique sentences makes long reasoning a direct consequence of the reward definition.

free parameters (5)

Reward weights αi, αt, αr, αv, αa = e.g., αa=0.80, αi=0.06, αt=0.06, αr=0.08 (Mixture/Continuous); αa=1.0 Only-Accuracy
Hand-tuned based on empirical observations (Table 6); no sensitivity analysis; determines which reward wins.
Continuous reward thresholds = numeric_diff<=5; hour_diff>2 => 0; minute_diff>10 => 0; decay denominators 5 and 20
Hand-set to define partial credit; no ablation; directly shapes continuous reward scores.
tanh smoothing / duplicate penalty = tanh(unique_sentences) - tanh(total-unique)
Arbitrary nonlinearity chosen to prevent reward hacking; affects relative benefit of longer reasoning.
GRPO hyperparameters = LR 1e-6, batch 128, KL 0.0, 500 steps, max lengths 2048
Chosen by the authors (Appendix A); no grid search; KL=0 unusual and not justified.
Textual representation encodings = character grids, coordinate lists, 'Current Time: 2:43'
Hand-designed per category; the bottleneck experiment's causal interpretation depends on these being exact and algorithmically neutral.

assumptions (6)

domain assumption AlgoPuzzleVQA gold test labels and question text are correct
Used as ground truth for all accuracy measurements; no independent verification by authors.
ad hoc to paper Hand-crafted text representations in Table 7 preserve all task-relevant information and do not reduce reasoning difficulty
This is the key identification assumption for the 'perception is bottleneck' claim; no control for representation ease.
domain assumption The modified data generator produces valid, unique puzzle instances without test-set overlap
Authors state uniqueness and non-overlap; no data released to check.
domain assumption GRPO implementation in EasyR1 is bug-free and behaves as expected
All RL results rely on this framework; no code/config released.
domain assumption Exact match on parsed <answer> tag is a valid measure of task accuracy
Open-ended answers parsed by a script; parsing errors could affect results.
domain assumption Single training run reflects expected performance
No seeds or repeats; multiple conclusions in Table 4 rest on small differences.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Disentangling Perception and Reasoning in Multimodal LLMs via Reward Design." pith.science (2026). https://pith.science/paper/XJHDCY3Y

@misc{pith2026260100215,
  author       = {Pith},
  title        = {Pith review of: Disentangling Perception and Reasoning in Multimodal LLMs via Reward Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XJHDCY3Y}},
  note         = {Machine review of arXiv:2601.00215}
}

read the original abstract

Reinforcement learning with verifiable rewards has driven major gains in LLM reasoning, and it is intuitive to assume this recipe will transfer well to multimodal models. However, multimodal models do two things: first, perceive what is in an image, then reason about what it implies. Because these stages are graded jointly, it is hard to tell how much room reasoning alone has to grow. We study this on algorithmic visual puzzles, where both components are necessary and show that perception, not reasoning, is the binding constraint. Replacing images with simple textual descriptions raises performance by over 20 points on average for Claude models. We then evaluate six reward designs aimed at inducing visual grounding during reasoning without chain-of-thought supervision. Training Qwen-2.5-VL-7B with GRPO, reward design induces long, structured reasoning with self-reflection and visual references, yielding a 5.56-point gain over the base model. These gains are, however, uneven; no single reward improves all categories, and rewards with verifiable accuracy signals trade out-of-domain transfer for in-domain accuracy. These results point to perception-aware reward design as a path forward, so that signals correct perception at its source rather than the reasoning that inherits its errors.

Figures

Figures reproduced from arXiv: 2601.00215 by the authors.

**Figure 2.** Generic prompt structure used across all cate [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** An example comparing reasoning between the base (Qwen-2.5-VL-7B) and RL-trained versions with [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (11 more)

**Figure 4.** Figure 4: Performance comparison across categories [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Question: Emily’s event is going to start in 1 hour 10 minutes. The current time is shown on the clock. The clock is a standard analog clock without the second hand. What time will the event start? Clock: This category features analog clocks with only hour and minute h…

**Figure 6.** Figure 6: Question: A checker game is being played on a grid of 4 squares with 1 green and 2 red checkers. Initially, the checkers are arranged as shown in the starting configuration, with the 3 checkers occupying 3 squares and one unoccupied square. Green checkers only move rig…

**Figure 8.** Figure 8: Question: A storekeeper is a puzzle in which the player pushes boxes around in a warehouse trying to get them to target locations. The game is represented by an 6 x 6 grid of characters grid where each element is a wall, floor, or box. Your task is to move the box to t…

**Figure 9.** Figure 9: Question: You are given a 9 * 9 chessboard. The Manhattan distance between two squares in a chessboard is equal to the minimal number of orthogonal King moves between these squares on the otherwise empty board. The objective is to place 9 chess queens on this board so …

**Figure 11.** Figure 11: Question: You are given a 3 x 3 grid in which each cell can contain either no kiwi, one fresh kiwi, or one rotten kiwi. Every minute, any fresh kiwi that is 4-directionally adjacent to a rotten kiwi also becomes rotten. What is the minimum number of minutes that must …

**Figure 10.** Figure 10: Question: The board shown in the image is a sliding puzzle of 3 * 3 tile dimensions. It has 8 numbered tiles and one unoccupied (open) position. Tiles in the same row or column of the open position can be moved by sliding them horizontally or vertically, respectively.…

**Figure 12.** Figure 12: Question: You are playing a Tower of Hanoi game with 3 rods and 6 disks of various diameters, which can slide onto any rod. You are given the starting and ending configuration of the game as shown in the top and the bottom of the image, respectively. The game has the …

**Figure 13.** Figure 13: Question: You are given 3 jugs of capacities 11, 5, 4 litres. Initially, the amount of water that is contained in each jar is shown in the image. A single step of water pouring from one jug to another is constrained by the following rules: i) take a non-empty jug and …

**Figure 14.** Figure 14: The RL-trained model generates detailed and well-structured reasoning for solving the puzzle, thoroughly [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]

**Figure 15.** Figure 15: The RL-trained model demonstrates structured multistep visual reasoning without any supervised training [PITH_FULL_IMAGE:figures/full_fig_p019_15.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Multimodal Mathematical Reasoning with Diverse Solving Perspective
cs.CL 2025-07 conditional novelty 6.0 of 10

Training a multimodal language model on multiple diverse solution paths per problem, plus rewards for distinguishing correct from incorrect solutions, improves math benchmark accuracy and output diversity.

Reference graph

Works this paper leans on

4 extracted references · 1 linked inside Pith · cited by 1 Pith paper

[1]

image description

is the top-left and (8, 8) is the bottom-right. The six queens are at: - (2, 3) – 2nd row, 3rd column - (3, 6) – 3rd row, 6th column - (4, 8) – 4th row, 8th column - (5, 1) – 5th row, 1st column - (6, 4) – 6th row, 4th column - (8, 7) – 8th row, 7th column To place the remaining two queens, I must find squares that are not in the same row, column, or diag...
[2023]

InThe Twelfth Inter- national Conference on Learning Representations

Let’s verify step by step. InThe Twelfth Inter- national Conference on Learning Representations. A.R. Luria. 1965. L.s. vygotsky and the problem of lo- calization of functions.Neuropsychologia, 3(4):387– 392. Brandon Ong, Tej Deep Pala, Vernon Toh, William Chandra Tjhi, and Soujanya Poria

1965
[2024]

thinking with images

How far are we from intelligent visual deduc- tive reasoning?arXiv preprint arXiv:2403.04732. Yaowei Zheng, Junting Lu, Shenzhi Wang, Zhangchi Feng, Dongdong Kuang, and Yuwen Xiong. 2025a. Easyr1: An efficient, scalable, multi-modality rl train- ing framework. https://github.com/hiyouga/ EasyR1. Ziwei Zheng, Michael Yang, Jack Hong, Chenxiao Zhao, Guohai ...

arXiv 2025
[2025]

Training vision-language process reward models for test-time scaling in multimodal reasoning: Key insights and lessons learned.arXiv preprint arXiv:2509.23250. OpenAI. 2024. Hello gpt-4o. https://openai.com/ index/hello-gpt-4o/. Rulin Shao, Shuyue Stella Li, Rui Xin, Scott Geng, Yiping Wang, Sewoong Oh, Simon Shaolei Du, Nathan Lambert, Sewon Min, Ranjay ...

arXiv 2024

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

image description

is the top-left and (8, 8) is the bottom-right. The six queens are at: - (2, 3) – 2nd row, 3rd column - (3, 6) – 3rd row, 6th column - (4, 8) – 4th row, 8th column - (5, 1) – 5th row, 1st column - (6, 4) – 6th row, 4th column - (8, 7) – 8th row, 7th column To place the remaining two queens, I must find squares that are not in the same row, column, or diag...

[2] [2023]

InThe Twelfth Inter- national Conference on Learning Representations

Let’s verify step by step. InThe Twelfth Inter- national Conference on Learning Representations. A.R. Luria. 1965. L.s. vygotsky and the problem of lo- calization of functions.Neuropsychologia, 3(4):387– 392. Brandon Ong, Tej Deep Pala, Vernon Toh, William Chandra Tjhi, and Soujanya Poria

1965

[3] [2024]

thinking with images

How far are we from intelligent visual deduc- tive reasoning?arXiv preprint arXiv:2403.04732. Yaowei Zheng, Junting Lu, Shenzhi Wang, Zhangchi Feng, Dongdong Kuang, and Yuwen Xiong. 2025a. Easyr1: An efficient, scalable, multi-modality rl train- ing framework. https://github.com/hiyouga/ EasyR1. Ziwei Zheng, Michael Yang, Jack Hong, Chenxiao Zhao, Guohai ...

arXiv 2025

[4] [2025]

Training vision-language process reward models for test-time scaling in multimodal reasoning: Key insights and lessons learned.arXiv preprint arXiv:2509.23250. OpenAI. 2024. Hello gpt-4o. https://openai.com/ index/hello-gpt-4o/. Rulin Shao, Shuyue Stella Li, Rui Xin, Scott Geng, Yiping Wang, Sewoong Oh, Simon Shaolei Du, Nathan Lambert, Sewon Min, Ranjay ...

arXiv 2024