PaperlayerView exact revision on arXiv

Source-linked AI summary

Same Reward, Different Skills: When Multimodal RL Learns to Look

Haocun Ye, Xinlong Jiang, Qile Chen, Bingyu Wang, Teng Zhang, Shubai Chen, Tingyu Wu, Zhenkun Zheng, Yiqiang Chen

arXiv:2610.01908v1cs.LG

TL;DR

RLVR benchmark gains are often read as perception improvements, but whether RLVR develops visual capabilities or makes existing ones easier to elicit remains open. The paper trains on counterfactual coordinate tasks designed to require visual evidence while remaining learnable, using standard GRPO and correctness-and-format rewards. On confirmatory 20-point scenes, a density absent from training, discovery accuracy rises from 0.425 to 0.875; gray canvases reduce it to zero, and matched blind training recovers essentially none.

  • Problem

    RLVR benchmark gains are often read as perception improvements, but whether RLVR develops visual capabilities or makes existing ones easier to elicit remains open.

    Source: §1 · p. 1, §8 · p. 9

  • Method

    The paper trains on counterfactual coordinate tasks designed to require visual evidence while remaining learnable, using standard GRPO and correctness-and-format rewards.

    Source: §1 · pp. 1–2

  • Results

    On confirmatory 20-point scenes, a density absent from training, discovery accuracy rises from 0.425 to 0.875; gray canvases reduce it to zero, and matched blind training recovers essentially none.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Task design determines what producing rewarded answers requires, so evaluation should track the intended visual operation rather than benchmark scores alone.

    Source: §8 · p. 9

    The evidence comes from coordinate scenes; generality to other visual formats and natural images, where visual necessity must first be measured, remains to be tested.

    Source: §8 · p. 9

Abstract

from arXiv · show

Reinforcement learning with verifiable rewards (RLVR) improves vision-language benchmark scores even without visual information during training. With images at test, blind-trained models recover roughly half of the real-image gain at 3B and nearly four fifths at 7B. Prolonged real-image training can erode grounding while benchmark gains persist. Both findings expose the same gap: an image in the prompt is not an image in the learning signal. Our design rule, visual resolvability, asks that visual evidence be necessary for a correct answer and that the task remain learnable. We test it on counterfactual coordinate scenes in which the question stays fixed and the target is never named, so a correct answer requires finding the target in the image. With standard GRPO and correctness-and-format rewards, a 7B model raises its accuracy at finding the target (discovery) from 0.425 to 0.875 on held-out scenes denser than any it trained on, and it improves on question types it never trained on. Two controls locate the source of the gain. Replacing test images with gray canvases drops discovery to zero; training on gray canvases instead, at matched step 30 and in each of four seeds, yields essentially none of the gain even when the model is then tested with real images. The learned skill carries over to grounding tasks built independently of the training corpus. A caption that answers the training question, added to the same images, reward and budget, cuts the gain by nearly two thirds. Changing what reward requires changes what RL learns.

1 INTRODUCTION

Benchmark gains can arise without visual information during training, while prolonged real-image training can erode grounding. The introduction proposes making visual evidence necessary for reward and tests that principle with relation-defined target tasks.

1 INTRODUCTION

  • Gray-image training recovers roughly half of real-image training gains at 3B and nearly four fifths at 7B, while prolonged real-image training can erode grounding despite benchmark gains.The introduction calls this loss grounding corrosion.Source: §1 · p. 1
  • Visual resolvability requires visual evidence for correct answers while keeping tasks learnable, making the intended visual operation part of earning reward rather than bypassable.The paper contrasts this with shortcut-accessible tasks that permit correct answers without visual evidence.Source: §1 · pp. 1–2
  • The study uses counterfactual scenes that hold questions fixed while changing a visual fact, with target cues ranging from marked locations or names to relations requiring discovery.It trains relation-only coordinate questions and target-identification probes with standard GRPO and correctness-and-format rewards.Source: §1 · p. 2
  • On a separate confirmatory set of 20-point scenes, a density absent from training, discovery accuracy rises from 0.425 to 0.875; gray-canvas testing reduces it to zero, while matched step-30 blind-trained controls evaluated with real images recover essentially none of the gain.Source: §1 · p. 2
  • The paper asks which visual operations improve and whether training images account for gains, holding rewards fixed within matched visual-access comparisons.Task construction specifies the evidence and operations successful answering requires.Source: §1 · p. 2

2 BENCHMARK GAINS ARE NOT VISUAL ACQUISITION

Benchmark gains from RLVR do not by themselves establish visual acquisition: blind training can recover gains, and ordinary training often favors reading supplied values over finding targets. The section separates training-time access, test-time image dependence, and the visual operation being evaluated.

2.1 LEARNING WITHOUT LOOKING

  • With real test images, REAL gains +0.238, NONE +0.137, and GRAY +0.125; blind training recovers 0.57 and 0.52 of REAL’s gain, respectively.Without test images, every condition gains little (+0.036 for REAL and +0.032 for GRAY).Source: §2.1 · p. 3, Table 1 · p. 3

2.2 HOW MUCH DOES THE IMAGE MATTER?

  • Chance-corrected blind retention separates BLINK from MathVerse at 0.27 versus 0.66, despite similar raw ratios of 0.83 and 0.85.The correction accounts for the chance reference for each item type.Source: §2.2 · p. 3

2.3 READING A VALUE, FINDING A TARGET

  • The cue hierarchy distinguishes marked-location readout, named-target find-and-bind, and relation-defined discovery, with each step supplying less target information.Counterfactual pairs hold the question fixed and change one visual fact so the correct answers differ.Source: §2.3 · p. 3
  • Cued readout gains +0.143 under REAL training, while REAL’s named-target grounding gain of +0.018 is equivalent to zero within the ±0.05 band.Figure 2(a) presents operation-level gains; its reported ordinary 3B discovery gains under blind conditions also finish within 0.04 of REAL at every density.Source: Figure 2 · p. 4, §2.3 · p. 4

3 WHEN GAINS CONCEAL GROUNDING LOSS

Under prolonged REAL training, benchmark accuracy rises while grounding-task accuracy falls below the base model’s. Separate gray-trained runs also lose overlapping previously solved grounding pairs, whereas the 7B access comparison shows a different balance.

3 WHEN GAINS CONCEAL GROUNDING LOSS

  • Benchmark accuracy rises from 0.175 to 0.436 and 0.453, while grounding pair accuracy falls by 0.070 and 0.052 to 0.407 and 0.413, below the base model’s 0.455.The decline occurs after a small rise during 400-step REAL training with a trainable vision encoder.Source: §3 · p. 4
  • Each of three GRAY-trained runs loses 45–51 previously solved grounding pairs; degraded sets overlap at Jaccard 0.661, and 39 shared wrong answers are identical across runs.The overlap exceeds a permutation-null mean of 0.011.Source: §3 · p. 4
  • At 7B, REAL gains +0.035 on the grounding task and +0.038 on its twin, while GRAY’s +0.012 and +0.013 changes have intervals including zero.Both models’ cued-readout changes also have intervals including zero.Source: §3 · p. 5

4 THE MISSING LINK: WHAT REWARD REQUIRES

The paper distinguishes needing an image from learning through visual evidence. It proposes tasks where visual evidence is necessary for correct answers and informative correctness feedback remains possible.

4 THE MISSING LINK: WHAT REWARD REQUIRES

  • A policy can exploit images at test without learning from them when nonvisual answer strategies suffice to earn reward.A target-location cue removes search while still requiring the value to be read from the image.Source: §4 · p. 5
  • A visually resolvable task ties correct answers to visual evidence and remains learnable, so successful solutions require the intended visual operation.The paper instantiates the rule in counterfactual coordinate scenes where necessity and learnability can be measured and controlled.Source: §4 · p. 5

5 MAKE LOOKING NECESSARY: MEASURE, SHAPE, VERIFY

The section operationalizes visual resolvability through an audit of visual benefit and learnability, then constructs counterfactual coordinate tasks where answers depend on changing visual facts and target cues. It verifies the test setting and specifies standard GRPO training.

5.1 MEASURE: VISUAL NECESSITY AND LEARNABILITY

  • The base-model audit estimates each image’s effect on correct-answer probability and checks whether prompts permit informative correctness feedback.The learnability band is 0 < b_i < 1; the constructed corpus has blind reward opportunity 0.003, and 92.9% of its rows meet both audit conditions.Source: §5.1 · p. 5

5.2 SHAPE: CHANGE THE FACT, CHANGE THE CUE

  • Counterfactual scene twins change one answer-relevant visual fact while keeping the question and remaining scene fixed.The K8 example moves the target while it remains the lowest point, changing the requested x-coordinate from −1 to 6.Source: §5.2 · p. 5
  • Discovery prompts supply only a relation, requiring target identification before coordinate readout, unlike cued readout and find-and-bind.The corpus includes target-stable, target-switch, and invariance pairs; 20-point scenes are reserved for evaluation, and L1/L2 are never trained.Source: §5.2 · p. 6

5.3 VERIFY: DOES THE TEST REQUIRE LOOKING?

  • The 20-point confirmatory scenes meet five preregistered criteria, including nontrivial learnability, zero image-removed discovery accuracy, and insufficient question-blind captions.The confirmatory scenes are among 450 new programs disjoint from training and development.Source: §5.3 · p. 6

5.4 TRAIN: STANDARD GRPO

  • Training uses standard GRPO with the vision encoder and projector frozen and the language model trainable.Each prompt receives five sampled responses; correctness and exact answer-span format are rewarded, with rewards normalized within each prompt’s group.Source: §5.4 · p. 6

6 THE REGIME FLIPS

Training on relation-only questions and identification probes improves discovery and untrained cued tasks on held-out scenes. The section tests visual access at evaluation and training, then measures transfer to independently built coordinate-grounding tasks.

6.1 LEARNING TO FIND WHAT TO READ

  • Discovery accuracy on the 20-point confirmatory set rises from 0.425 to 0.740 after 30 steps and 0.875 after 100, a gain of 0.450.At step 100, untrained cued readout rises from 0.610 to 0.890 and find-and-bind from 0.620 to 0.950.Source: §6.1 · p. 7
  • Training reward exceeds 0.97 by step 13, while discovery on 12- and 20-point development scenes continues rising through step 75.A second run ends within 0.010 of the first at every development density; three seeds reach 0.730–0.760 on the confirmatory set at step 30.Source: §6.1 · p. 7
  • Training find-and-bind instead of relation-only questions leaves confirmatory discovery +0.042 lower, while separately trained identification and grounding compose into discovery.On base-model-unidentified members, REAL lifts discovery from 0.143 to 0.629; GRAY improves neither group.Source: §6.1 · p. 7

6.2 USING IMAGES, LEARNING FROM IMAGES

  • Removing test images reduces discovery and probe accuracy to zero; with real images at test, matched-step gray-canvas training recovers essentially none of REAL’s gain.Across four seeds, REAL exceeds GRAY by +0.354 on the confirmatory set; adding answer-bearing captions yields +0.120 versus REAL’s +0.323 discovery gain.Source: §6.2 · pp. 7–8

6.3 DISCOVERY TRAINING TRANSFERS TO COORDINATE GROUNDING

  • At step 100, grounding-task pair accuracy rises from 0.768 to 0.948, and its independently regenerated twin rises from 0.728 to 0.937.At matched step 30, REAL gains +0.162 on grounding while the GRAY-trained control gains +0.002.Source: §6.3 · p. 8

7 TURNING THE DIAL

The section varies the constructed-data share at a fixed training budget and compares discovery learning curves across model sizes. A one-third constructed mixture delivers most of the full-corpus discovery gain, while performance differs by scale and density.

7.1 CHANGING THE MIXTURE, CHANGING THE GAIN

  • The one-third mixture matches the untrained model on ViRL39K (0.519 against 0.509), MathVista testmini (0.696 against 0.682), and BLINK (0.547 against 0.536).Ordinary-only training reaches 0.548 on ViRL39K, while the fully constructed run reaches 0.419.Source: §7.1 · p. 8
  • At equal constructed exposure, the one-third mixture reaches 0.745 against 0.675 for the fully constructed run.Both conditions receive 2,400 constructed-data presentations.Source: §7.1 · p. 8
  • On never-trained 20-point scenes, discovery gain is −0.010 with ordinary data, +0.275 with a one-third constructed mixture, and +0.330 with the full corpus.The one-third mixture delivers over 80% of the full-corpus gain at the same budget; the full corpus gives the largest gain at every density.Source: §7.1 · p. 8, Figure 5 · p. 9

7.2 SAME TASKS, DIFFERENT LEARNING CURVES

  • On held-out training-density scenes, 3B at step 30 roughly matches 7B at step 10, but on the never-trained density reaches 0.545 against 0.675.Grounding-task performance improves at both scales: +0.130 for 3B and +0.162 for 7B at step 30.Source: Figure 5 · p. 9, §7.2 · p. 9

8 DISCUSSION

The discussion argues that task design determines what a fixed verifier rewards learners to acquire, so evaluation and checkpointing should track the intended skill rather than aggregate score alone.

8 DISCUSSION

  • With verifier, model, reward, optimizer, and budget fixed, changing task composition changes acquisition, making task design a determinant of what RL learns.The discussion frames the learning problem as the lever, with data construction one way to make the intended operation necessary for success.Source: §8 · p. 9
  • Benchmark improvement, dependence on visual input, and learning from visual evidence are distinct claims that require different tests.Image removal tests dependence on the image; matched blind training tests whether improvement was learned from visual evidence.Source: §8 · p. 9
  • Checkpoint selection should track the skill being taught because benchmark gains can persist while grounding erodes.The discussion also notes that discovery can keep improving after reward saturates.Source: §8 · p. 9
  • The evidence is scoped to coordinate scenes; testing other visual formats and natural images, where visual necessity must first be measured, remains future work.Within the coordinate testbed, training also improves untrained question types and independently constructed grounding tasks.Source: §8 · p. 9
  • The experiments keep the vision encoder and projector frozen, while the broader proposal ties verifiable success to a chosen evidence source and a learnable operation.The discussion names acoustic distinctions, temporal order, and tool-returned results as possible evidence requirements beyond vision.Source: §8 · pp. 9–10

REPRODUCIBILITY STATEMENT

  • The repository provides data, training recipes, evaluation outputs, and code to rebuild most reported tables and figure values without a GPU.Re-running training and evaluation requires GPUs; six printed values are not rebuilt by the script.Source: REPRODUCIBILITY STATEMENT · p. 10

A VISUAL NECESSITY AUDIT

  • The visual-necessity audit reports image-present and blind accuracy, chance, and naive and chance-corrected retention, with further audits of Gemma-3 and InternVL3-9B.For Gemma-3, corrected retention is undefined when image-present accuracy falls below chance; InternVL3-9B's ratio is unstable when its margin over chance is small.Source: A VISUAL NECESSITY AUDIT · p. 13, Table A.1 · p. 14, Table A.2 · p. 14

B TASK CONSTRUCTION

The task construction uses counterfactual coordinate scenes and held-out scene programs to test discovery, while a confirmatory evaluation checks whether the intended visual task is learnable and not shortcut-accessible.

B TASK CONSTRUCTION

  • Each scene program yields twin images differing in one fact, with three question levels that supply progressively less target information and a relation-identification probe.The corpus includes target-stable, target-switch, and invariance pair roles in equal numbers.Source: B TASK CONSTRUCTION · p. 14, Table B.1 · p. 15
  • Discovery is scored by member accuracy on target-stable and invariance pairs, while target-switch pairs, cued levels, and probes are reported separately.The independently constructed grounding suite uses pair accuracy.Source: B TASK CONSTRUCTION · p. 14
  • Training uses 2,880 prompts from 720 scene programs with 8 or 12 points; L1 and L2 prompts and 20-point scenes are reserved from training.Training, development, and confirmatory scenes come from disjoint scene programs.Source: B TASK CONSTRUCTION · p. 14
  • Only the 20-point confirmatory cell meets all five evaluation criteria; it supplies the confirmatory result.The criteria include bounded discovery accuracy, lower discovery than cued levels, zero gray-canvas and no-image accuracy, a caption-only threshold, and an artifact screen.Source: B TASK CONSTRUCTION · p. 15, Table B.4 · p. 16
  • 72.1% of REAL’s pair-count-weighted improvement across the 1,200-pair suite comes from cued readout.Hierarchy discovery, measured by member accuracy, is excluded from this weighting.Source: B TASK CONSTRUCTION · p. 16

C INDEPENDENTLY CONSTRUCTED GROUNDING EVALUATIONS

The independently constructed grounding evaluations use counterfactual pairs built before the training corpus and validate that image evidence matters; the regenerated twin is evaluated only for 7B models.

C INDEPENDENTLY CONSTRUCTED GROUNDING EVALUATIONS

  • The 1,200-pair suite contains three independently constructed grounding tasks whose paired renders differ in one answer-relevant visual fact.The tasks were built before the training corpus and share no scenes with it.Source: C INDEPENDENTLY CONSTRUCTED GROUNDING… · p. 16, Table C.1 · p. 16
  • Validation finds at most 0.054 caption-only pair accuracy for Qwen2.5-VL-72B-Instruct and exactly 0.000 pair accuracy without images across the listed conditions.Human review confirmed all 60 audited pairs; Gemma-3 and InternVL3-9B also score exactly 0.000 without images.Source: Table C.1 · p. 16
  • The independently regenerated twin was evaluated only for the 7B models.Source: C INDEPENDENTLY CONSTRUCTED GROUNDING… · p. 16

D TRAINING AND CANONICAL ACCESS RESULTS

The appendix documents the training and matched visual-access comparisons, then reports grounding corrosion under prolonged training and clarifies how gains and visual attribution are measured. The reported results include a 7B blind-training recovery share and an extractability analysis of blind gains.

D TRAINING AND CANONICAL ACCESS RESULTS

  • The 30-step matched comparisons ended before any training restart; the vision encoder and projector were frozen in constructed-corpus runs.The 100-step REAL runs used restarts at steps 30, 50, and 75, resetting optimizer moments and re-anchoring the KL reference.Source: D TRAINING AND CANONICAL ACCESS RESULTS · p. 17
  • The GRAY-trained control’s mean accuracy reward rose only from 0.0008 at step 1 to 0.031 at step 30, while its format reward rose from 0.87 to 1.00.Across three further seeds, GRAY accuracy reward at step 30 was 0.021–0.048, versus 0.972–0.991 for REAL and named-target runs.Source: D TRAINING AND CANONICAL ACCESS RESULTS · p. 17

E CORROSION

  • Grounding pair accuracy fell from step 100 to step 400 by 0.070 and 0.052, ending 0.048 and 0.042 below the base model.Benchmark accuracy peaked at step 200; the runs changed horizon, encoder, corpus filter, and reward implementation together, so they do not isolate a cause.Source: E CORROSION · p. 17, Table E.1 · p. 21

D TRAINING AND CANONICAL ACCESS RESULTS

  • GRAY recovered 0.779 [0.642, 0.921] of REAL’s real-tested gain on Geometry3K at 7B.The table reports the recovery share as GRAY’s gain divided by REAL’s.Source: Table D.4 · p. 19

F GAINS, ATTRIBUTION, AND MIXTURE MEASUREMENT

  • Among 497 questions with an extractable untrained real-image answer, accuracy was 0.211; the split shows blind gains were not confined to questions lacking such answers.The passage explicitly cautions that this descriptive split does not establish how much of any gain is due to format.Source: F GAINS, ATTRIBUTION, AND MIXTURE… · p. 20

G ACQUISITION, ATTRIBUTION, AND TRANSFER

The section evaluates acquisition on held-out scenes, compares real-image training with gray-trained controls, and tests transfer and ablations. REAL-trained models improve discovery and transfer, while gray-test performance and the matched gray-training control help distinguish visual access at evaluation from learning through images.

G ACQUISITION, ATTRIBUTION, AND TRANSFER

H DOSE, AUDIT STANDARDIZATION, AND CAPACITY

The appendix examines how constructed-data mixtures relate to acquisition, describes scoring and uncertainty conventions, and analyzes paired evaluation and correctness feedback. Its results distinguish fixed-budget mixture outcomes from equal-exposure comparisons and characterize when reward groups contain informative correctness variation.

H DOSE, AUDIT STANDARDIZATION, AND CAPACITY

  • On the never-trained density, the two intermediate mixtures were tied at +0.275 and +0.270, while the fully constructed condition gave the largest gain on every density.On the fixed-budget grid, audited mass m(f) was nearly linear in f, with correlation 0.997 and largest residual 0.006.Source: H DOSE, AUDIT STANDARDIZATION, AND… · p. 27, Table H.1 · p. 27
  • On held-out ViRL39K, real-image accuracy was 0.548, 0.519, 0.522, and 0.419 for f = 0, 1/3, 2/3, and 1.The fully constructed run’s deficit was concentrated on multiple-choice items; the one-third mixture matched the untrained model on MathVista testmini and BLINK, while the fully constructed run fell on both.Source: H DOSE, AUDIT STANDARDIZATION, AND… · p. 27

I SCORING AND UNCERTAINTY

  • Scoring uses the same answer matcher for trained and untrained models, and grounding-suite accuracy requires both members of a pair to be correct.Intervals resample the task’s evaluation unit with trained models held fixed; constructed-task units are scene programs, while grounding pairs and benchmark items are resampled as pairs or items.Source: I SCORING AND UNCERTAINTY · pp. 27–28

J.1 ANSWER-CHANGING PAIRED EVALUATION

  • An image-independent deterministic predictor has zero pair accuracy on answer-changing pairs when nonvisual inputs match and accepted-answer sets are disjoint.This proposition concerns pair accuracy, not member accuracy on discovery tasks or the target-stable identification probe.Source: §J.1 · p. 30

J.2 AUDIT AND CORRECTNESS FEEDBACK

  • At step 1, 58–67% of real-image discovery groups contained both correct and incorrect responses, compared with 2–5% for gray-canvas groups.Gray-canvas groups were 95–100% all wrong; later, format variation diminished and updates depended on the rare group containing a correct response.Source: §J.2 · p. 30, Table J.1 · p. 31

J.3 COMPARING AUDIT MEASUREMENTS UNDER A COMMON SUCCESS PROFILE

  • The standardized audit places measurements on a shared initial-success profile; it changes the audit summary, not the training distribution.The appendix separately reports associations and equal-exposure comparisons.Source: §J.3 · p. 31

K THE EXPERIMENTAL PROGRAM AT A GLANCE

The experimental program surveys RLVR across task families, corpora, benchmarks, models, and training runs, while extended related work situates its controls and task design among prior approaches. Its reported results span unseen question types, public-corpus mixtures, and models, alongside comparisons of what visual training and reward requirements contribute.

K THE EXPERIMENTAL PROGRAM AT A GLANCE

  • At step 30, final-answer pair accuracy rises from 19 to 33 of 80 pairs at 5 points and from 22 to 28 at 20 points on never-trained premise questions.On the grounding task and its twin, +0.160 (95% CI [0.130, 0.190]) of the gain is attributable to training images, versus +0.002 for the GRAY-trained control.Source: K THE EXPERIMENTAL PROGRAM AT A GLANCE · p. 32
  • On ViRL39K’s 4,239 held-out items, REAL gains +0.204 and GRAY gains +0.149; with constructed-data mixing, never-trained discovery gains +0.275 versus −0.010 for all-ViRL39K training.Every mixture’s held-out ViRL39K accuracy is at least 0.20 above its image-removed accuracy.Source: K THE EXPERIMENTAL PROGRAM AT A GLANCE · p. 32
  • Discovery improves at both trained scales, while two untrained model families score 0.333 and 0.681 on grounding and fall to 0.000 without images.The program spans nine counterfactual task families, two public training corpora, seven audited benchmarks, four models from three families, and over thirty GRPO runs; Figure K.1 shows eight task families.Source: K THE EXPERIMENTAL PROGRAM AT A GLANCE · p. 32, Figure K.1 · p. 32

L EXTENDED RELATED WORK

  • Prior RLVR studies leave open whether training develops skills or elicits them; this work compares blind and image-informed training while holding verifier and setup fixed within corpora.Related work also reports grounding corrosion under prolonged real-image training with a trainable vision encoder, even as benchmark gains persist.Source: L EXTENDED RELATED WORK · p. 33
  • Visual-dependence benchmarks test shortcuts and image reliance, while the counterfactual suite changes one visual fact under a fixed question and scores paired answers.Because accepted answers are disjoint, an image-independent greedy predictor scores exactly zero on pair accuracy; discovery is scored per question and measured directly without images.Source: L EXTENDED RELATED WORK · p. 33
  • Prior perception-aware methods alter objectives, response structure, evidence, rewards, or tasks; these experiments use standard GRPO with correctness-and-format rewards fixed within constructed-corpus comparisons.Matched blind-trained controls use the same prompts, starting checkpoint, reward, and budget to estimate the contribution of training-time visual information.Source: L EXTENDED RELATED WORK · pp. 33–34
  • The learnability band describes corpus learnability without filtering, while resolvability also asks whether visual evidence and a required operation determine successful answering.The paper contrasts its sampled-correctness measure of image benefit with difficulty-based selection and VisNec’s teacher-forced loss measure.Source: L EXTENDED RELATED WORK · p. 34
  • ViCrit and Visual Jigsaw build verifiable perception tasks, while this work adds matched blind-training controls, a cue hierarchy, and shortcut-accessible twins.The comparison also distinguishes its sampled-correctness change under RL verification from VisNec’s teacher-forced loss reduction.Source: L EXTENDED RELATED WORK · p. 34
Loading 2610.01908v1…