Pith.

Computer Vision and Pattern Recognition · arXiv:2610.16679 v2 · submitted 12 Oct 2026 · 163 references

EvoSpatial: Experience-Driven Skill and Tool Evolution for Spatial Reasoning Agents

Bin Yu, Shijie Lian, Zhaolong Shen, Xiaopeng Lin, Yao Zhang, Haishan Liu, Haotian Li, Yuliang Wei

show 5 more authors

Kai Hu, Yichao Du, Zhirui Zhang, Bailing Wang, Kai Chen

The paper says

“Spatial reasoning agents can inspect tool outputs and repair errors, but reusing these lessons requires preserving both the computation and the conditions that make it valid.”

more of the abstract

We introduce EvoSpatial, a framework that learns from execution experience by updating a library of linked skills and tools. Skills describe procedures, geometric assumptions, and verification steps; tools implement the operations those procedures require. When solving a task, the agent retrieves relevant skills and synthesizes reusable computational functions if existing tools are insufficient. After each task, execution feedback and evidence checks guide skill revision and tool consolidation, without ground-truth answers or benchmark scores. Across 20 benchmarks, EvoSpatial averages 63.9, 4.0 percentage points above the average reported by SpatialClaw. The evolved library also improves held-out and cross-benchmark performance when updates are disabled, and trajectories generated with evolved skills improve supervised fine-tuning, even without external skills at evaluation. Keeping checked procedures on record, together with the tools that support them, lets an agent carry spatial reasoning experience from one task to the next.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 15 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. A serious empirical systems paper whose frozen-library and transfer results mostly support the reuse claim; send to review, but the same-backbone judge and missing uncertainty reporting keep the mechanism from being fully established. This line is wrong

2 major objections · 6 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 163 references say.

The paper says, at Section 3.3, Appendix A, Appendix D (Table 9), Section 4.4 (Table 5)

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.67):

“These assessments guide library management but do not establish causal benefit.”

Pith says: Objection 1 of 2 · major

The LLM judge is load-bearing: skill admission, promotion, revision, suspension, and the reliability statistics entering Eq. (3) all come from 'separate sessions of the configured backbone' with no independently trained evaluator (Appendix A). The only quantitative evidence of judge reliability is Table 9's 85.3% human agreement 'on the audited records', but the manuscript does not describe the sample size, the selection of audited cases, the human rater protocol, or how 'unsupported acceptance' is defined and scored.

more of objection 1

The authors do acknowledge that these assessments 'do not establish causal benefit' (Appendix A), and the held-out gains in Table 3 do rule out the naive version of self-confirmation, because held-out scores are computed against benchmark labels with updates disabled. What remains is an attribution gap: every condition in Table 5 retains the same-backbone judge, and the append-only variant (63.7) already recovers more than half of the full system's gain over the initial library (62.0 to 65.4), so no ablation isolates judge filtering from procedure accumulation. Please describe the audit completely, report agreement broken down by verdict type and by whether the judged entry was generated by the judging backbone itself, and add at least one condition in which admission and utility assessment use an independent evaluator or purely deterministic checks.

The paper says, at Table 2 (note), Section 4.2, Tables 7-8

The review did not quote this passage.

Pith says: Objection 2 of 2 · major

The headline 4.0-point gain over SpatialClaw compares EvoSpatial's three-run means against baseline scores quoted from the SpatialClaw paper ('Baseline scores are reported by SpatialClaw (Cho et al., 2026)') rather than measured under the manuscript's protocol (768px longer-edge resizing, 32-frame sampling, 30 tool-call turns, fixed-seed 1,000-example sampling for large benchmarks).

more of objection 2

Protocol choices affect both systems, so the reported gap is not necessarily the gap under matched conditions, and it is not stated which benchmarks exceed 1,000 examples and are therefore subsampled. The same concern applies to Tables 7 and 8, where SpatialClaw token and turn counts are presented without saying whether they are re-runs under the described kernel and budgets or numbers from the original paper. Please state this explicitly and provide matched-protocol re-runs of SpatialClaw for at least the 10 benchmarks in Table 3, ideally all 20, so the central comparison does not rest on cross-paper protocol assumptions.

6 minor objections

The paper says, at Abstract; Section 4.2; Appendix B

Pith says: minor objection 1 of 6

The abstract and conclusion headline the 63.9 online-evolution number, which is measured while the library adapts on the same benchmark questions being scored; Appendix B discloses this clearly, but I recommend presenting the fixed-library held-out result (Table 3) as the primary generalization claim in the abstract, since the two numbers measure different quantities.

The paper says, at Table 2, Average row

Pith says: minor objection 2 of 6

There is a formatting error in the Average row ('59.963.9' with a missing space), and the 'VADAR — *' entries would be clearer if the asterisk were explained in the caption rather than only at the end of the row.

show 4 more

The paper says, at Tables 3 and 5

Pith says: minor objection 3 of 6

The held-out and ablation results report only three-run means, and the differences among full (65.4), append-only (63.7), and no-guidance (64.1) conditions are in the 1.4-1.8 point range; reporting run-level variance or paired differences would show that these gaps are not within-run noise.

The paper says, at Appendix A, Eq. (3) and lifecycle rules

Pith says: minor objection 4 of 6

The ranking weights (lambda_a = 0.05, lambda_r = 0.02), the archival penalty, the capacity limits (64 skills, 16 helpers), and the promotion threshold (0.70) are hand-set, and no sensitivity analysis is given; a small parameter sweep would increase confidence that the Table 5 findings are not threshold artifacts.

The paper says, at Section 4.5, Figure 5

Pith says: minor objection 5 of 6

The cross-benchmark transfer means (+1.73 and +1.76 over the initial library) are reported without standard deviations or run-level detail; adding those would make the transfer claim comparable in rigor to the rest of the paper.

The paper says, at References

Pith says: minor objection 6 of 6

Several references lack complete metadata (for example, the 'Gemma 4' entry gives only a URL and access date, and 'Zhang et al., 2025a' has no venue), and the first entry of the BLINK reference contains a line-break typo ('Blink: Mul- 9 timodal'); please clean up the bibliography.

What would settle it

Re-run the Section 4.3 held-out protocol (10 benchmarks, 20%/80% split, three runs each) with the LLM judge's support/contradict/inconclusive verdicts replaced by (i) random verdicts at the same marginal rates and (ii) an independent-backbone judge (e.g., Qwen3.8-27B), keeping retrieval, reflection, curation, and budgets fixed. If arm (i) collapses toward the 62.0 initial-library score while the reported 65.4 persists under arm (ii), the same-backbone judge's verdicts carry genuine utility information and the attribution holds. If arm (i) also reaches about 65.4, judge-dependent filtering is not the source of the gains and the experience-driven attribution must be revised toward procedure accumulation.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 15 Oct 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic partly agrees 2 min 18 s · 20,888 tokens read, 17,998 written
  3. Editor wrote the verdict above 24 s · 21,890 tokens read, 2,867 written
  4. Referee major revision: 2 major objections, 6 minor 1 min 36 s · 21,265 tokens read, 12,443 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

EvoSpatial is a genuine empirical systems paper, not a repackaging exercise. It transplants the self-evolving skill-library loop from coding and web agents into spatial reasoning, and it does the work needed to make that transplant informative: skills explicitly carry geometric assumptions and verification steps, tools are linked to skills and consolidated, and the evaluation separates online evolution from frozen-library reuse. The fixed-library result is the paper's best evidence: evolve on 20% of ten benchmarks, then evaluate on the remaining 80% with updates off, and the joint library moves the average from 62.0 to 65.4. Cross-benchmark transfer from ERQA and MindCube also looks plausible, and the SFT experiment is cleanly matched. Credit is due for three-run means, same-split handling for shared visual inputs, matched budgets in ablations, and a limitations section that is candid rather than decorative.

The headline 4.0-point gain over SpatialClaw is the weakest part of the evidence, because it compares against SpatialClaw's reported numbers rather than reruns under matched conditions. That is a minor-to-moderate issue: the paper's own fixed-library and ablation results do not depend on those numbers.

The bigger soft spot is the LLM judge. Section 3.3 and Appendix A make clear that selection, reflection, curation, and utility assessment all use separate sessions of the same backbone, and Appendix D's 85.3% human agreement is reported without describing the audit sample or rater protocol. This matters because the judge is the gate for admission, promotion, and revision. The fixed-library experiments do defuse the naive version of the worry: a self-confirming judge cannot fabricate ground-truth held-out gains. But the attribution gap remains: no condition in Table 5 runs without judge-dependent filtering, and the append-only variant still lands at 63.7, so roughly half the held-out gain survives without curation. The mechanistic claim that checked procedures, rather than sheer accumulated volume, drive the gain is not fully isolated.

Two smaller points: Table 3 and Table 5 report means over three runs without error bars or significance tests; skill-only and tool-only variants differ from joint by only 1.6 and 1.4 points, which could be noise. And each backbone evolves its own fresh library, so cross-backbone generality of the reusable library is untested. Neither sinks the paper. The citation pattern looks normal for this area.

Verdict: this deserves a serious referee. I would send it to review and ask for code and data, uncertainty reporting on the fixed-library numbers, and either an independent evaluator or a properly documented audit before counting the mechanism fully established.

The strongest objection

The central claim — that checked spatial experience, not merely added procedure volume, drives reuse — rests on the LLM judge separating supported from contradicted applications. Section 3.3 assigns the judge this load-bearing role, and Appendix A states that selection, reflection, curation, and utility assessment all use 'separate sessions of the configured backbone' with no independently trained evaluator; the paper itself flags that these assessments 'do not establish causal benefit' and that a tool call completing without error does not establish a correct spatial inference ([b91], [b336], [b337]). The only direct evidence that verdicts are informative is Appendix D's 85.3% human agreement 'on the audited records', with sampling, sample size, and rater protocol undescribed. The fixed-library experiments (Section 4.3) do rule out the naive version: a purely self-confirming judge cannot fabricate ground-truth held-out gains, and this is real credit to the design. The residual soft spot is attribution: every ablation in Table 5 still uses the same-backbone judge for admission and usage statistics, so no condition exists without judge-dependent filtering, and the append-only variant (63.7 vs 65.4) shows that nearly half of the held-out gain survives without curation. If the judge systematically favors entries its own backbone generated, ground-truth gains can arise from procedure accumulation and retrieval scaffolding while the mechanistic claim — and any cross-backbone generality of the 'reusable' library, which Section 4.7 never tests since each backbone evolves its own fresh library — remains unestablished.

I read the paper as making a reuse claim: spatial reasoning experience can be retained as a linked skill-tool library and reused across tasks. For reuse to be the right explanation of the reported gains, the judge's support/contradict verdicts must carry information about genuine spatial utility, because those verdicts are the only filtering component for admission, promotion, revision, consolidation, and the reliability statistics in the retrieval scoring function. The paper's Appendix A is explicit that no independently trained evaluator is used and that assessments do not establish causal benefit; Appendix D reports 85.3% agreement on 'audited records' without describing the audit. That is the least secure condition in the argument. I did not make the SpatialClaw baseline comparison the primary concern: baseline scores are quoted from a cited source, and the reuse claim is independently supported by within-paper controlled experiments (Tables 3, 5, 6; Figure 5) that do not depend on SpatialClaw's numbers. Those within-paper experiments are the strongest evidence, and I credit the fixed-library protocol: with updates disabled and scores computed against benchmark labels, a self-confirming judge cannot manufacture the held-out gains, and the consistent cross-benchmark and SFT improvements corroborate genuine library utility. The remaining, more precise soft spot is attribution. Table 5's ablations control task-specific guidance, verification guidance, and curation, but no condition removes judge-dependent filtering; the append-only condition (63.7) still admits candidates through the same-judge pipeline and reaches nearly half of the full gain (65.4). If the judge is lenient toward its own backbone's candidates, gains of this shape would also arise from procedure accumulation without the judge carrying utility information, leaving the central mechanism claim unverified and the library's transferability across backbones untested (each row of Table 4 evolves its own library). This does not warrant rejection — the direction of the effects is well supported — but it does warrant conditional acceptance with the counterfactual check above, which is where the reader's verdict and mine coincide. I therefore keep CONDITIONAL (UNCHANGED) and mark partial agreement with the reader: the root concern is the same, but I read the fixed-library design as already defusing the strongest self-confirmation version, so the failure mode I take most seriously is the attribution of gains to judge-driven evolution and the untested cross-backbone generality of the library, rather than shrinkage of the headline numbers under an independent evaluator.

The referee's report

The controlled protocol is stronger than is typical for this literature: three independent runs with perturbed question order, 20/80 evolution/evaluation splits with questions sharing visual inputs kept in the same split, matched budgets for ablations, and a fixed-library regime in which the evolved library is evaluated with updates and helper synthesis disabled against ground-truth labels. These choices rule out the naive self-confirmation scenario: a judge that merely approved its own outputs could not, by itself, produce the held-out gains in Table 3, and that is genuine credit to the design. The consistent per-benchmark advantage over SpatialClaw in Table 2 and the token-efficiency results are also useful reference points. The remaining weakness is attribution: every condition in the ablations retains the same-backbone judge, so the mechanistic claim that checked and curated experience, rather than accumulated procedure volume plus retrieval scaffolding, drives reuse is not isolated as an experimental factor.

Pith's reading of the paper (a machine's summary, not the authors')

EvoSpatial sets out to show that a spatial reasoning agent can carry experience from one task to the next if the lessons are kept as a linked library of skills and tools, rather than spent inside a single answer. Its central claim is that this retention works: across 20 benchmarks the framework averages 63.9, 4.0 points above the average SpatialClaw reports, and the gains are not merely continued adaptation, since a library evolved on part of the data and then frozen still lifts held-out performance from 62.0 to 65.4. Because the library is built from execution feedback and evidence checks, no ground-truth answers or benchmark scores enter the update loop, and the same library that improves reasoning also feeds supervised fine-tuning. The point that matters, if the paper is right, is that a spatial agent's checked procedures are a reusable asset — the computation together with the geometric conditions that justify it.

What carries the argument. The load-bearing object is the persistent skill–tool library and the update loop that maintains it. A skill is a recorded procedure with conditions of use, geometric assumptions, and verification steps; a tool is a callable operation, either a perception model or a generated helper; every skill is explicitly linked to the tools it uses, so a retrieved skill tells the agent not only what to compute but what to check before trusting the result. During a task, the Skill Selector retrieves complementary skills and emits task-specific usage guidance, and the Tool Synthesizer writes a deterministic helper when existing tools are insufficient. After the task, the Experience Reflector extracts only lessons supported by checked operations or observed repairs, and the Library Curator applies local edits and merges overlapping helpers. The pair of connected mechanisms — retrieval with guidance, and curation gated by the LLM judge's supported/contradicted/inconclusive verdicts — is what turns single-task repairs into reusable procedures, by keeping the computation and the condition of its validity together.

Core claim. On the paper's own terms, the discovery is that the reusable unit of spatial experience is not a tool call or a repair but a skill–tool pair: a skill records a multi-step procedure, its geometric assumptions, and the verification steps that must pass, while linked tools supply the computations. EvoSpatial alternates between solving tasks and updating this library, synthesizing new helper functions when existing tools cannot perform a needed operation, and after each task reflecting on execution records and evidence checks to revise or consolidate entries. An LLM judge classifies each observed application as supported, contradicted, or inconclusive, and those verdicts drive admission, promotion, revision, and archival. On 20 benchmarks with the main backbone, the system averages 63.9 against SpatialClaw's reported average of 59.9, with the largest category gains in multi-view (+4.6) and video spatial/4D (+5.7) reasoning; each benchmark starts from a fresh copy of the initial library, so the headline gain comes from within-benchmark online evolution, with cross-benchmark carryover tested separately. When updates are disabled, jointly evolved skills and tools raise the held-out average from 62.0 to 65.4, beating either component alone; libraries evolved on ERQA or MindCube improve the 18 common target benchmarks by 1.73 and 1.76 points respectively; and supervised fine-tuning on skill-guided trajectories beats fine-tuning on unguided trajectories by 1.2 to 1.4 points in every task group, even though skills are absent at evaluation.

The premise it rests on. The load-bearing premise is that the LLM judge's supported/contradicted/inconclusive verdicts — produced by the same backbone that generated the skills — are accurate enough to drive which entries are admitted, promoted, revised, or merged; the paper reports 85.3% agreement with human judgments on audited records but does not describe the audit, so a systematically self-confirming judge would make the library an artifact rather than genuinely reusable experience.

If the paper is right (Pith's inferences, not the authors')

Spatial reasoning agents can improve across tasks without any ground-truth answers or benchmark scores: execution feedback and local evidence checks are the only signals the update loop consumes.

Experience survives as a portable asset: a library frozen after evolution on part of a benchmark still raises held-out accuracy, and libraries evolved on one benchmark improve other benchmarks with different task structure.

The largest gains concentrate where evidence must be combined across views or time (multi-view +4.6, video spatial/4D +5.7 category points), suggesting the library's main value is in verification-heavy reasoning rather than quick perceptual judgment.

Evolved skills change the training signal too: supervised fine-tuning on trajectories generated with skills beats fine-tuning on matched unguided trajectories by 1.2–1.4 points in every task group, even with skills removed at evaluation.

Experience reuse also pays in cost: on ERQA the evolved library raises accuracy while cutting total tokens by 20.9% and mean tool-call turns from 5.09 to 4.32 with the main backbone.

If judge-gated curation is as reliable as the reported 85.3% human agreement suggests, the same skill–tool loop should transfer to other perception-grounded domains — robot manipulation, navigation, or inspection — where outcomes are locally checkable but global labels are scarce; the paper demonstrates only vision benchmarks.

Because the frozen library transfers across benchmarks and the gains hold across five different backbones, nothing in the mechanism ties a library to one model; a testable next step is exchanging libraries between agents with different backbones, which the paper does not run.

The self-confirmation risk is directly measurable: compare held-out gains from helpers and skills promoted at the judge's 0.70 confidence threshold against those promoted at lower confidence; if verdicts were noise, the correlation would be flat.

The BLINK pattern — every tool-augmented agent trails the no-tool baseline on intuitive perceptual questions — suggests the library's value is task-dependent; a reasonable extension is a router that skips procedure following for rapid-judgment questions, reaping the multi-view gains without the hallucination cost the paper cites.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper · Review history (2 versions)

The paper

Abstract

Spatial reasoning agents can inspect tool outputs and repair errors, but reusing these lessons requires preserving both the computation and the conditions that make it valid. We introduce EvoSpatial, a framework that learns from execution experience by updating a library of linked skills and tools. Skills describe procedures, geometric assumptions, and verification steps; tools implement the operations those procedures require. When solving a task, the agent retrieves relevant skills and synthesizes reusable computational functions if existing tools are insufficient. After each task, execution feedback and evidence checks guide skill revision and tool consolidation, without ground-truth answers or benchmark scores. Across 20 benchmarks, EvoSpatial averages 63.9, 4.0 percentage points above the average reported by SpatialClaw. The evolved library also improves held-out and cross-benchmark performance when updates are disabled, and trajectories generated with evolved skills improve supervised fine-tuning, even without external skills at evaluation. Keeping checked procedures on record, together with the tools that support them, lets an agent carry spatial reasoning experience from one task to the next.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-14

The authors' main claim, as the classifier picked it out 0.98

“These results support accumulating spatial experience for both inference-time reuse and trajectory-based training.”

How the main claim is stated: asserted 0.72

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.12 · human participants 0.07 · controlled comparison 0.87 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.04

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Tool-augmented vision-language agents compose perception and measurement to reason about spatial relationships, but effective execution also requires tracking geometric assumptions and checking intermediate evidence. background 0.90 0.08 0.03 asserted 1.00
Preserving this experience across tasks requires retaining procedural knowledge together with the computations it governs. background 0.67 0.08 0.05 asserted 0.97
We introduce EvoSpatial, an experience-driven framework that evolves a library of linked skills and tools. contribution 1.00 0.89 0.07 asserted 1.00
Our key insight is that successful and failed spatial trajectories reveal both how to compute and when to trust the result. contribution 0.55 0.56 0.18 asserted 1.00
EvoSpatial retrieves skills, reuses tools, and synthesizes reusable computational functions (generated helpers) during reasoning. method 0.98 0.76 0.07 asserted 1.00
After answer submission, supervised outcome feedback guides skill refinement and tool consolidation for subsequent tasks. method 0.99 0.81 0.07 asserted 1.00
Across 20 benchmarks, EvoSpatial attains the highest average score among the compared spatial agents, exceeding SpatialClaw by 4.0 percentage points. result 1.00 0.96 0.22 asserted 1.00
Further experiments show cross-benchmark transfer and improved reasoning through skill-guided distillation without external skills at evaluation. result 1.00 0.95 0.13 asserted 1.00
These results support accumulating spatial experience for both inference-time reuse and trajectory-based training. interpretation 1.00 0.73 0.69 asserted 0.72

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.97

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-14: 1 request, 6,730 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (5)
Figure 1 of the paper
Figure 1. Interactive spatial reasoning across views. Tools support correspondence, frame alignment, and distance comparison. et al., 2026). The agent must then choose com￾patible operations, check their assumptions, and determine whether the measurements support an answer. Existing spatial agents learn tool coordination or compose operations through visual programs and interactive execution (Chen et al., 2026; Luo et al., 2026; Cho et al., 2026). Interactive execution allows an agent to inspect intermediate results and repair errors before answering. A repaired error can also provide a les
Figure 2 of the paper
Figure 2. Tools and skills in EvoSpatial. Tools provide callable operations; skills specify how to combine them and check the resulting evidence. For each task, EvoSpatial selects relevant skills and generates task-specific usage guidance. When existing tools cannot perform a required computa￾tion, it synthesizes a helper. The agent executes code, inspects the results, and repairs errors be￾fore submitting an answer. Afterwards, it extracts reusable lessons from checked operations and ob￾served repairs, revises skills, and consolidates gen￾erated helpers; the updated library is then available…
Figure 3 of the paper
Figure 3. Overview of EvoSpatial. (a) Resident and retrieved skills guide interactive spatial reasoning with linked tools. (b) Execution feedback guides helper synthesis, reflection, and library updates for subsequent tasks. et al., 2026; Lian et al., 2026b; Tan et al., 2026; Lian et al., 2026a) provide geometric represen￾tations that complement direct VLM predictions. SpaceTools learns spatial tool coordination through interactive training (Chen et al., 2026), while pyS￾patial composes spatial operations in visual pro￾grams (Luo et al., 2026). SpatialClaw uses multi￾turn execution to invoke…
Figure 4 of the paper
Figure 4. From failure reflection to skill reuse. A revised contact-verification skill directs the agent to inspect local surfaces before answering in a subsequent scene. 0 +1 +2 +3 +4 Gain over Initial (percentage points) Single-image Multi-view Video / 4D General spatial General video Omni3D OmniSpatial SPBench MMSI SPAR-Bench MMSI-Video OSI-Bench PAI-Bench VSI-Bench-U VSTI-Bench DSI-Bench BLINK SpatialTree ViewSpatial CV-Bench PerceptComp Video-MME Video-MME-v2 Cross-benchmark transfer ERQA-evolved (Avg. +1.73) MindCube-evolved (Avg. +1.76)
Figure 5 of the paper
Figure 5. Cross-benchmark transfer from ERQA and MindCube. Score gains over the initial library on 18 common target benchmarks, excluding both sources. Libraries remain fixed during target evaluation. Training condition Single￾image Multi￾view General Avg. (9) Base Model 58.6 54.7 62.4 58.6 Unguided Trace SFT 59.6 55.9 63.2 59.6 Skill-Guided Trace SFT 60.8 57.3 64.4 60.9

Figures are the authors', reproduced from the paper.

References (163)
  1. [1] Vggt: Visual geometry grounded transformer , author= on Pith
  2. [2] OpenVLA: An Open-Source Vision-Language-Action Model , author= on Pith
  3. [3] Black, Kevin and Brown, Noah and Driess, Danny and Esmail, Adnan and Equi, Michael and Finn, Chelsea and Fusai, Niccolo and Groom, Lachy and Hausman, Karol and Ichter, Brian and others , journal= on Pith
  4. [4] Physical Intelligence and Kevin Black and Noah Brown and James Darpinian and Karan Dhabalia and Danny Driess and Adnan Esmail and Michael Equi and Chelsea Finn and Niccolo Fusai and Manuel Y. Galliker and Dibya Ghosh and Lachy Groom and Karol Hausman and Brian Ichter and Szymo... on PithPith's reviewarXiv
  5. [5] GR00T N1: An Open Foundation Model for Generalist Humanoid Robots , author=. 2025 , eprint= on Pith
  6. [6] Actions as Language: Fine-Tuning VLMs into VLAs Without Catastrophic Forgetting , author=. 2025 , eprint= on Pith
  7. [7] VLM4VLA: Revisiting Vision-Language-Models in Vision-Language-Action Models , author=. 2026 , eprint= on Pith
  8. [8] ChatVLA: Unified Multimodal Understanding and Robot Control with Vision-Language-Action Model , author=. 2025 on Pith
  9. [9] DualVLA: Building a Generalizable Embodied Agent via Partial Decoupling of Reasoning and Action , author=. 2025 , eprint= on Pith
  10. [10] Peebles, William and Xie, Saining , title =. 2023 , pages = on Pith
  11. [11] Visual Instruction Tuning , author= on Pith
  12. [12] Qwen2.5-VL Technical Report , author=. arXiv preprint arXiv:2502.13923 , year= on PithPith's reviewarXiv
  13. [13] Qwen3-VL Technical Report , author=. arXiv preprint arXiv:2511.21631 , year= on PithPith's reviewarXiv
  14. [14] Evaluating Real-World Robot Manipulation Policies in Simulation , author= on Pith
  15. [15] RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots , author=. Robotics: Science and Systems , year= on Pith
  16. [16] StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developing , author =. 2025 , month =. doi:10.5281/zenodo.18264214 , howpublished = on Pithdoi:10.5281/zenodo.18264214
  17. [17] Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration , author=. 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2024 , organization= on Pith
  18. [18] Octo: An Open-Source Generalist Robot Policy , author=. 2024 , eprint= on Pith
  19. [19] Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success , author=. 2025 , eprint= on Pith
  20. [20] Towards Generalist Robot Policies: What Matters in Building Vision-Language-Action Models , author=. 2024 , eprint= on Pith
show 80 more
  1. [21] TraceVLA: Visual Trace Prompting Enhances Spatial-Temporal Awareness for Generalist Robotic Policies , author=. 2025 , eprint= on Pith
  2. [22] SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Model , author=. 2025 , eprint= on Pith
  3. [23] CogACT: A Foundational Vision-Language-Action Model for Synergizing Cognition and Action in Robotic Manipulation , author=. 2024 , eprint= on Pith
  4. [24] VideoVLA: Video Generators Can Be Generalizable Robot Manipulators , author=. 2025 , eprint= on Pith
  5. [25] FAST: Efficient Action Tokenization for Vision-Language-Action Models , author=. 2025 , eprint= on Pith
  6. [26] GR00T N1.6: An Improved Open Foundation Model for Generalist Humanoid Robots , author =. 2025 , month = on Pith
  7. [27] Universal Actions for Enhanced Embodied Foundation Models , author =. 2025 , pages = on Pith
  8. [28] TinyVLA: Towards Fast, Data-Efficient Vision-Language-Action Models for Robotic Manipulation , author=. 2025 , eprint= on Pith
  9. [29] SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics , author=. 2025 , eprint= on Pith
  10. [30] What Can RL Bring to VLA Generalization? An Empirical Study , author=. 2026 , eprint= on Pith
  11. [31] X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model , author=. 2025 , eprint= on Pith
  12. [32] RynnVLA-002: A Unified Vision-Language-Action and World Model , author=. 2025 , eprint= on Pith
  13. [33] MM-ACT: Learn from Multimodal Parallel Generation to Act , author=. 2025 , eprint= on Pith
  14. [34] InternVLA-A1: Unifying Understanding, Generation and Action for Robotic Manipulation , author=. 2026 , eprint= on Pith
  15. [35] Evo-1: Lightweight Vision-Language-Action Model with Preserved Semantic Alignment , author=. 2025 , eprint= on Pith
  16. [36] SpatialTree: How Spatial Abilities Branch Out in MLLMs , author=. 2026 , eprint= on Pith
  17. [37] LLaVA-3D: A Simple yet Effective Pathway to Empowering LMMs with 3D-awareness , author=. 2025 , eprint= on Pith
  18. [38] DSI-Bench: A Benchmark for Dynamic Spatial Intelligence , author=. 2025 , eprint= on Pith
  19. [39] RynnEC: Bringing MLLMs into Embodied World , author=. 2025 , eprint= on Pith
  20. [40] Visual Spatial Tuning , author=. 2025 , eprint= on Pith
  21. [41] Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning of Vision Language Models , author=. 2025 , eprint= on Pith
  22. [42] RoboRefer: Towards Spatial Referring with Reasoning in Vision-Language Models for Robotics , author=. 2026 , eprint= on Pith
  23. [43] Shiqi Chen and Tongyao Zhu and Ruochen Zhou and Jinghan Zhang and Siyang Gao and Juan Carlos Niebles and Mor Geva and Junxian He and Jiajun Wu and Manling Li , booktitle=ICML, year=. Why Is Spatial Reasoning Hard for on Pith
  24. [44] Spatial Forcing: Implicit Spatial Representation Alignment for Vision-language-action Model , author=. 2025 , eprint= on Pith
  25. [45] Cambrian-S: Towards Spatial Supersensing in Video , author=. 2025 , eprint= on Pith
  26. [46] Jiang, Jingjing and Ma, Chao and Song, Xurui and Zhang, Hanwang and Luo, Jun , title =. 2025 , pages = on Pith
  27. [47] RoboBrain 2.0 Technical Report , author=. 2025 , eprint= on Pith
  28. [48] MiMo-Embodied: X-Embodied Foundation Model Technical Report , author=. 2025 , eprint= on Pith
  29. [49] Galaxea Open-World Dataset and G0 Dual-System VLA Model , author=. 2025 , eprint= on Pith
  30. [50] XR-1: Towards Versatile Vision-Language-Action Models via Learning Unified Vision-Motion Representations , author=. 2025 , eprint= on Pith
  31. [51] Igniting VLMs toward the Embodied Space , author=. 2025 , eprint= on Pith
  32. [52] Hybrid Training for Vision-Language-Action Models , author=. 2025 , eprint= on Pith
  33. [53] InstructVLA: Vision-Language-Action Instruction Tuning from Understanding to Manipulation , author=. 2025 , eprint= on Pith
  34. [54] Robotic Control via Embodied Chain-of-Thought Reasoning , author=. 2025 , eprint= on Pith
  35. [55] DeepThinkVLA: Enhancing Reasoning Capability of Vision-Language-Action Models , author=. 2025 , eprint= on Pith
  36. [56] OneTwoVLA: A Unified Vision-Language-Action Model with Adaptive Reasoning , author=. 2025 , eprint= on Pith
  37. [57] EO-1: Interleaved Vision-Text-Action Pretraining for General Robot Control , author=. 2025 , eprint= on Pith
  38. [58] VLA-R1: Enhancing Reasoning in Vision-Language-Action Models , author=. 2025 , eprint= on Pith
  39. [59] VLA-Reasoner: Empowering Vision-Language-Action Models with Reasoning via Online Monte Carlo Tree Search , author=. 2025 , eprint= on Pith
  40. [60] ThinkAct: Vision-Language-Action Reasoning via Reinforced Visual Latent Planning , author=. 2025 , eprint= on Pith
  41. [61] Fast-ThinkAct: Efficient Vision-Language-Action Reasoning via Verbalizable Latent Planning , author=. 2026 , eprint= on Pith
  42. [62] MolmoAct: Action Reasoning Models that can Reason in Space , author=. 2025 , eprint= on Pith
  43. [63] Don't Blind Your VLA: Aligning Visual Representations for OOD Generalization , author=. 2025 , eprint= on Pith
  44. [64] How Do VLAs Effectively Inherit from VLMs? , author=. 2025 , eprint= on Pith
  45. [65] Rajbhandari, Samyam and Rasley, Jeff and Ruwase, Olatunji and He, Yuxiong , title =. Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis , articleno =. 2020 , isbn = on Pith
  46. [66] Decoupled Weight Decay Regularization , author=. 2019 , eprint= on Pith
  47. [67] PhysBrain: Human Egocentric Data as a Bridge from Vision Language Models to Physical Intelligence , author=. 2025 , eprint= on Pith
  48. [68] Evaluating Object Hallucination in Large Vision-Language Models , author=. 2023 , eprint= on Pith
  49. [69] Seed-bench: Benchmarking multimodal llms with generative comprehension , author=. arXiv preprint arXiv:2307.16125 , year= on PithPith's reviewarXiv
  50. [70] Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering , author=. The 36th Conference on Neural Information Processing Systems (NeurIPS) , year= on Pith
  51. [71] Chaoyou Fu and Peixian Chen and Yunhang Shen and Yulei Qin and Mengdan Zhang and Xu Lin and Jinrui Yang and Xiawu Zheng and Ke Li and Xing Sun and Yunsheng Wu and Rongrong Ji and Caifeng Shan and Ran He , booktitle= on Pith
  52. [72] LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning , author=. arXiv preprint arXiv:2306.03310 , year= on PithPith's reviewarXiv
  53. [73] Lin, Yixin and Wang, Austin S. and Sutanto, Giovanni and Rai, Akshara and Meier, Franziska , title = on Pith
  54. [74] Mixture-of-Transformers: A Sparse and Scalable Architecture for Multi-Modal Foundation Models , author=. Transactions on Machine Learning Research , issn=. 2025 , note= on Pith
  55. [75] CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models , author=. 2025 , eprint= on Pith
  56. [76] LangForce: Bayesian Decomposition of Vision Language Action Models via Latent Action Queries , author=. 2026 , eprint= on Pith
  57. [77] Limited Linguistic Diversity in Embodied AI Datasets , author=. 2026 , eprint= on Pith
  58. [78] ABot-M0: VLA Foundation Model for Robotic Manipulation with Action Manifold Learning , author=. 2026 , eprint= on Pith
  59. [79] TwinBrainVLA: Unleashing the Potential of Generalist VLMs for Embodied Tasks via Asymmetric Mixture-of-Transformers , author=. 2026 , eprint= on Pith
  60. [80] Evo-0: Vision-language-action model with implicit spatial understanding , author=. 2025 , eprint= on Pith
  61. [81] Wang, Jianyuan and Chen, Minghao and Karaev, Nikita and Vedaldi, Andrea and Rupprecht, Christian and Novotny, David , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2025 , pages = on Pith
  62. [82] Training Verifiers to Solve Math Word Problems , author=. 2021 , eprint= on Pith
  63. [83] Measuring Massive Multitask Language Understanding , author=. International Conference on Learning Representations , year= on Pith
  64. [84] When Vision Overrides Language: Evaluating and Mitigating Counterfactual Failures in VLAs , author=. 2026 , eprint= on Pith
  65. [85] A New Era of Intelligence with Gemini 3 , year = on Pith
  66. [86] AgileX Robotics , title = on Pith
  67. [87] Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware , author=. 2023 , eprint= on Pith
  68. [88] Diffusion Policy: Visuomotor Policy Learning via Action Diffusion , author=. 2024 , eprint= on Pith
  69. [89] Liu, Songming and Wu, Lingxuan and Li, Bangguo and Tan, Hengkai and Chen, Huayu and Wang, Zhengyi and Xu, Ke and Su, Hang and Zhu, Jun , booktitle =. RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation , volume = on Pith
  70. [90] Open-sourcing GO-1: The Bitter Lessons of Building VLA Systems at Scale , author =. 2025 , month = on Pith
  71. [91] DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control , author=. arXiv preprint arXiv:2502.05855 , year= on PithPith's reviewarXiv
  72. [92] Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation , author=. IEEE Robotics and Automation Letters (RA-L) , year= on Pith
  73. [93] Accelerating Vision-Language-Action Model Integrated with Action Chunking via Parallel Decoding , author=. arXiv preprint arXiv:2503.02310 , year= on PitharXiv
  74. [94] StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developing , author=. arXiv preprint arXiv:2604.05014 , year= on PithPith's reviewarXiv
  75. [95] PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models , author=. 2026 , eprint= on Pith
  76. [96] UAM: A Dual-Stream Perspective on Forgetting in VLA Training , author=. 2026 , eprint= on Pith
  77. [97] PaLM-E: An Embodied Multimodal Language Model , author=. 2023 , eprint= on Pith
  78. [98] RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control , author=. 2023 , eprint= on Pith
  79. [99] Oier Mees and Lukas Hermann and Erick Rosete-Beas and Wolfram Burgard , title =. IEEE Robotics and Automation Letters (RA-L) , volume= on Pith
  80. [100] VLABench: A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks , author=. 2024 , eprint= on Pith

The first 100 of 163 references are listed here. All of them are in this paper's citation record.

Cited by 2 Pith papers

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment