Pith.

Computer Vision and Pattern Recognition · arXiv:2611.24578 v1 · submitted 18 Nov 2026 · 84 references

BaseRSI: Recursive Self-Improvement of a Base Vision-Language Model via Training-Free Harness Internalization

Mingda Zhang, Yuanhao Dong, Qiang Huang, Rui Qian, Yanjin Li, Zhenyu Liu, Xiao Huang, Zihao Yang

show 3 more authors

Jian Wang, Pinjia He, Xiaoying Tang

The paper says

“Recursive self-improvement (RSI) is moving from the harness of prompts and tools around a base model toward the model itself.”

more of the abstract

However, existing RSI still faces key challenges, including harness-bound gains that stay outside the base model, black-box supervision that reads only the model's outputs, and imprecise edits whose site and strength are fixed in advance. To address these challenges, we propose BaseRSI, a new paradigm of Recursive Harness Internalization: the base model improves its harness, then writes what the harness adds into its own weights through training-free edits, round after round on its own responses. To support this paradigm, we introduce ReGap, a mechanism that supervises every transformer block with the high-dimensional activation difference that the harness makes at that block alone, and updates the block's weights along this signal in a low-dimensional subspace. Building on this mechanism, we further propose ReKey, an algorithm that writes this signal by regression into a few key-value neurons of every block, and measures the site and strength of each edit on the model's own responses. Moreover, the base model fills every role of this recursion with forward passes alone, and its edits, written at different sites, compose in one checkpoint without refitting. Experimental results on ten visual reasoning benchmarks show that BaseRSI consistently improves a base vision-language model both in distribution and on held-out benchmarks. Our code is available at https://github.com/beita6969/BaseRSI.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 19 Nov 2026 · read the full text (195,116 characters) · not checked by a person

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

Holds up. BaseRSI: Recursive Self-Improvement of a Base Vision-Language Model via Training-Free Harness Internalization This line is wrong

11 minor

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

Are you an author? What you can do.

The paper, first character to last

The reviewer was given the whole text, 195,116 characters. Not checked: figures, code, what the 84 references say.

11 minor objections

The paper says, at §5, Table 1, main results

Pith says: minor objection 1 of 11

The main results (Table 1) show that BaseRSI outperforms all baselines on every benchmark. However, the comparison with fine-tuning methods (SFT, GRPO) may not be entirely fair: these methods are trained on the development questions until convergence, but the paper does not specify whether they use the same token budget or reasoning effort during inference as BaseRSI.

more of objection 1

The caption states 'all methods but the two closed-source reference models use Qwen3.8-27B with the same reasoning effort, token budget, and evaluator', but the fine-tuning methods change the weights, so the comparison of inference cost is not directly relevant. More importantly, the fine-tuning baselines are trained on the same development questions that BaseRSI uses for its acceptance test and fitting, so they have access to the same data. The fact that they barely improve suggests that the gain comes from the harness internalization paradigm, but the paper should discuss whether the fine-tuning methods could benefit from the same harness entries (e.g., by training on the teacher's responses).

The paper says, at §4.2, Eq. (6)

Pith says: minor objection 2 of 11

The notation ω_β(z) = z σ(β z) is defined but the sigmoid function σ is not explicitly defined. It should be stated that σ is the logistic sigmoid.

show 9 more

The paper says, at §4.3, Eq. (8)

Pith says: minor objection 3 of 11

The ridge regression in Eq. (8) uses a mean gap b_l and a projection matrix P_l. The definition of P_l as 'removes the mean direction of the inputs' is clear, but the text does not specify how the ridge weight λ is chosen. Appendix A.4 states that λ is chosen on a held-out fold, but this detail should be mentioned in the main text for completeness.

The paper says, at §5, Table 1

Pith says: minor objection 4 of 11

The table caption says 'mean_{±std} over three runs'. The standard deviations are reported, but the number of runs (3) is small for reliable variance estimation. The paper should note this limitation, especially for benchmarks with high variance (e.g., BabyVision std 2.4 for BaseRSI).

The paper says, at §5, Fig. 4

Pith says: minor objection 5 of 11

Figure 4 panels are described in the caption but the actual figures are not visible in the manuscript text. The captions are detailed enough to understand the results, but the paper should ensure that the figures are legible in the final version.

The paper says, at Appendix A.4, Algorithm 3

Pith says: minor objection 6 of 11

Algorithm 3 (ReKey, one block) states 'Rescale each readout to unit variance and its value by the inverse factor'. This step is important for numerical stability but is not explained in the main text. A brief justification would be helpful.

The paper says, at §4.1, Eq. (3)

Pith says: minor objection 7 of 11

The acceptance test uses a one-sided sign-flip p-value. The description 'obtained by flipping the signs of the image clusters inside each benchmark' is clear, but the paper should specify how many flips are used and whether the test is exact or approximate.

The paper says, at §5, Table 2

Pith says: minor objection 8 of 11

Table 2 (ablations) reports accuracy for many variants. The row 'random readouts' and 'random values' are described as null edits. The paper should clarify whether these null edits are applied on top of the view edit (as stated in the caption) or as standalone edits.

The paper says, at §5, Fig. 5

Pith says: minor objection 9 of 11

Figure 5 panels (c)–(f) show mechanism checks. The caption mentions 'Inside vs. output' and 'Exactness of the gap', but the axes and units are not described in the caption. The paper should ensure that all figure elements are self-contained.

The paper says, at §4.3, Eq. (9)

Pith says: minor objection 10 of 11

The key fitting in Eq. (9) uses three groups of positions. The paper should clarify whether the key is fitted once per block or once per entry (across all blocks). The algorithm description suggests one key per block, but the text is ambiguous.

The paper says, at §5, Table 1

Pith says: minor objection 11 of 11

The baseline 'LBS-27B' is described as a released checkpoint fine-tuned from the same base model. The paper should cite the source of this checkpoint and specify whether it was fine-tuned on the same development questions or on different data.

What would settle it

No significant objection identified. The paper's claims are well-supported by the evidence provided.

from the skeptic's reading

How the review went

5 passes by deepseek-v4-flash · 19 Nov 2026

  1. Reader holds up confidence high
  2. Skeptic agrees 5 s · 68,843 tokens read, 197 written
  3. Referee 11 minor 16 s · 69,225 tokens read, 2,411 written
  4. Check against the paper kept 0 of 4 objections 1 min 52 s · 70,014 tokens read, 21,728 written
  5. Editor wrote the verdict above 6 s · 69,439 tokens read, 379 written

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

Pith's passes raised 4 objections. A check against the paper's text kept 0, withdrew 3 (3 Pith's reviewer misread the paper) and moved 1 to the minor objections.

The review in full

The review in brief, machine-written

This paper presents a method for a vision-language model to improve itself by writing what its own harness (prompts, tools, image views) adds into its weights, using forward passes only. The core idea is to treat the activation difference that a harness entry makes at each transformer block as a target, and to fit a low-rank linear readout of each block's MLP input to that gap via ridge regression, then write the result into a few key-value neurons. The method is evaluated on ten visual reasoning benchmarks (five in-distribution, five held-out) using Qwen3.8-27B, and it consistently outperforms prompting, visual zoom-in, harness evolution, and fine-tuning. The gains transfer to other base models (Gemma-4-31B, Qwen3.6-27B, Qwen3.5-9B). The paper is thorough: it provides mathematical propositions with proofs in the appendix, extensive ablations, statistical tests, and released code. The main claims are supported by the evidence. The soft spots are minor. The proofs assume Lipschitz constants and bounded pairs that are not explicitly verified, but the empirical results are consistent with the theory. The rank-64 readout is justified empirically (Table 12) but not theoretically guaranteed. The acceptance test assumes that gains on validation questions generalize to held-out benchmarks, which is a standard assumption and is tested. The paper is well-written and the experiments are carefully designed. This paper deserves a serious referee. It is a solid contribution to the field of recursive self-improvement and weight editing. I recommend engaging with the work.

The strongest objection

No significant objection identified. The paper presents a well-motivated method with clear theoretical foundations, extensive experiments, and thorough ablations. The claims are supported by the evidence. The method is novel and the results are significant. The paper is well-written and the code is provided. No critical red flags were identified.

The paper presents a well-motivated method with clear theoretical foundations, extensive experiments, and thorough ablations. The claims are supported by the evidence. The method is novel and the results are significant. The paper is well-written and the code is provided. No critical red flags were identified.

The referee's report

Not published. Pith's referee wrote this assessment before its objections were checked against the paper, and 3 of its 4 major objections did not survive that check.

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

The paper proposes BaseRSI, a method for a vision-language model to improve its own weights without training, by first evolving a 'harness' (prompts/tools) that helps it answer questions, then writing the effect of that harness into its own weights via a training-free edit. The key idea is that the harness's effect on each transformer block's activations (the 'gap') can be measured and then regressed into a few key-value neurons per block, composing edits from different harness entries. The paper claims this consistently improves the base model on all tested benchmarks, outperforming prompting, fine-tuning, and other harness-evolution methods, and that the edited model exceeds its own harness in the request while answering original requests alone.

What carries the argument. The central mechanism is the teacher-student gap (ReGap): for each transformer block, the difference in the block's output when the harness is present vs. absent, computed on the same response states. This gap is the exact descent direction that would make the student match the teacher at that block. ReKey then writes this gap into the model's weights by fitting a low-rank linear readout of the MLP input via ridge regression, storing the result in a few key-value neurons per block (r=64). The site of the edit (which weights to change) is determined by measuring which part of the divergence from the teacher is removed by patching the linear-attention layers vs. the MLP. The strength (step size) is set by a line search on held-out responses. The outer loop evolves the harness by proposing candidate entries from the model's own failures and accepting those whose gain generalizes across benchmarks (leave-one-benchmark-out significance test).

Core claim. The paper claims that a base vision-language model can recursively improve its own weights by internalizing the gains from its harness (prompts, tools, image views) through a training-free mechanism called ReGap and ReKey. ReGap measures the per-block activation difference that the harness makes (the teacher-student gap) and treats it as a high-dimensional signal. ReKey writes this signal into a few key-value neurons per block via ridge regression, with the site and strength of each edit measured on the model's own responses. The edits from different harness entries compose in one checkpoint without refitting. The paper reports that this method consistently improves the base model on ten visual reasoning benchmarks (five in-distribution, five held-out), outperforming prompting, visual zoom-in, harness evolution, and fine-tuning, and that the edited checkpoint exceeds its own harness in the request while answering original requests on standard inference engines.

The premise it rests on. The method assumes that the teacher-student activation gap at each block can be faithfully approximated by a low-rank (r=64) linear readout of the MLP input, and that the acceptance test (leave-one-benchmark-out significance) ensures generalization to held-out benchmarks.

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

If BaseRSI works as claimed, a vision-language model can improve its own weights on its own responses without any training (backward passes, gradient updates, or external supervision), using only forward passes and linear solves.

The method would allow a base model to consolidate gains from multiple harness entries (e.g., a finer image view and a counting procedure) into a single checkpoint, where the edits compose without refitting and each can be removed individually.

The recursive loop (outer: evolve harness, inner: internalize into weights) could in principle continue indefinitely, with the model proposing new entries from its improved state and internalizing them in subsequent rounds.

Because the edits are written into a small fraction of weights (0.23% of the language model) and leave the rest bit-identical, the method is interpretable and reversible, and the edited model can be deployed on standard inference engines without the harness in the request.

The paper's approach suggests a general paradigm for self-improvement that could extend beyond vision-language models to any transformer-based architecture, as long as a 'harness' (augmented input) can be defined and its per-block activation gap measured.

The reliance on a low-rank (r=64) linear readout of the MLP input to approximate the teacher-student gap implies that the harness's effect on each block is largely linear in the MLP input; if this assumption fails for more complex harness entries, the method might require higher rank or non-linear readouts.

The acceptance test (leave-one-benchmark-out significance) is a practical way to ensure generalization, but it depends on the representativeness of the validation split; for deployment in truly novel domains, the test might need to be adapted or supplemented with additional validation data.

The paper's results show the edited model sometimes exceeds the teacher (harness in the request) on out-of-scope questions, which the authors attribute to the student learning with privileged information (the key that reads its own state). This suggests that internalization can provide benefits beyond simply copying the harness's behavior, potentially leading to better generalization.

Check this review

What was read

The passes were given all 195,116 characters of the paper's LaTeX source.

Fingerprint of that text (SHA-256): 7e3ca86e40402f1e32cc86a77cb1867fcc7b5e02f5145458c9a8ffc405f4857a

Pith does not keep that text. Rebuilt from the paper's source or PDF by the same code, it has this fingerprint if it is the same text. From a copy of Pith's code:

PYTHONPATH=. python scripts/review_input_check.py 2611.24578

The passes, in the order they ran

  1. Reader deepseek-v4-flash · medium effort · $0.016 · instructions of the Reader, fingerprint f9ead7300382…
  2. Skeptic deepseek-v4-flash · medium effort · $0.0063 · instructions of the Skeptic, fingerprint 8f88c0f61488…
  3. Referee deepseek-v4-flash · medium effort · $0.0067 · instructions of the Referee, fingerprint a852d99aea0d…
  4. Check against the paper deepseek-v4-flash · medium effort · $0.016 · instructions of the Check against the paper, fingerprint 6c7a6883d551…
  5. Editor deepseek-v4-flash · medium effort · $0.0064 · instructions of the Editor, fingerprint e7211f6a3d27…
  6. Circularity check deepseek-v4-flash · medium effort · $0.010 · instructions of the Circularity check, fingerprint e9946e39d67b…
  7. Summary for the page deepseek-v4-flash · medium effort · temperature 0.2 · $0.0086 · instructions of the Summary for the page, fingerprint e4f56bf568e7…

The passes cost $0.070 in model calls together.

Version stamped on this review: pith-pipeline@v1.4.1-daily-deepseek-or-medium+deepseek/deepseek-v4-flash+p.d60f54226f00

Each link opens a pass's instructions as plain text, the exact bytes its fingerprint was taken over. Every stored set of instructions.

curl -s https://pith.science/methodology/prompts/f9ead730038216984cfb56d7cb34332bdf691828296a1de4245f3d6667d4bf56 | sha256sum

prints the fingerprint of the reader's instructions. The same works for every pass.

The receipt

The review's fingerprint was taken on 19 Nov 2026. The review you are reading still matches it.

Fingerprint of the review (SHA-256): 2977841aeae544201f377e7a13a127b144ae100533e08f0ca8b98a20e5c03cd1

curl -s 'https://pith.science/api/v1/verdicts/2611.24578?form=canonical' | sha256sum

prints the fingerprint of the review as it stands now. The rule, in four steps.

A matching fingerprint shows the review has not changed since that date. It does not show that the review is right.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, from the paper's LaTeX source.

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

The paper

Abstract

Recursive self-improvement (RSI) is moving from the harness of prompts and tools around a base model toward the model itself. However, existing RSI still faces key challenges, including harness-bound gains that stay outside the base model, black-box supervision that reads only the model's outputs, and imprecise edits whose site and strength are fixed in advance. To address these challenges, we propose BaseRSI, a new paradigm of Recursive Harness Internalization: the base model improves its harness, then writes what the harness adds into its own weights through training-free edits, round after round on its own responses. To support this paradigm, we introduce ReGap, a mechanism that supervises every transformer block with the high-dimensional activation difference that the harness makes at that block alone, and updates the block's weights along this signal in a low-dimensional subspace. Building on this mechanism, we further propose ReKey, an algorithm that writes this signal by regression into a few key-value neurons of every block, and measures the site and strength of each edit on the model's own responses. Moreover, the base model fills every role of this recursion with forward passes alone, and its edits, written at different sites, compose in one checkpoint without refitting. Experimental results on ten visual reasoning benchmarks show that BaseRSI consistently improves a base vision-language model both in distribution and on held-out benchmarks. Our code is available at https://github.com/beita6969/BaseRSI.

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 · 2026-11-19

The authors' main claim, as the classifier picked it out 1.00

“Experimental results on ten visual reasoning benchmarks show that BaseRSI consistently improves a base vision-language model both in distribution and on held-out benchmarks.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.11 · states limitation 0.13 · human participants 0.03 · controlled comparison 0.28 · randomized assignment 0.01 · states quantitative result 0.38 · compares with prior results 0.30 · mentions code or data availability 0.99

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
Recursive self-improvement (RSI) is moving from the harness of prompts and tools around a base model toward the model itself. background 1.00 0.06 0.07 asserted 0.99
However, existing RSI still faces key challenges, including harness-bound gains that stay outside the base model, black-box supervision that reads only the model's outputs, and imprecise edits whose site and strength are fixed in advance. gap 0.77 0.07 0.03 asserted 1.00
To address these challenges, we propose BaseRSI, a new paradigm of Recursive Harness Internalization: the base model improves its harness, then writes what the harness adds into its own weights through training-free edits, round after round on its own responses. contribution 0.96 0.47 0.20 asserted 0.98
To support this paradigm, we introduce ReGap, a mechanism that supervises every transformer block with the high-dimensional activation difference that the harness makes at that block alone, and updates the block's weights along this signal in a low-dimensional subspace. method 0.86 0.82 0.05 asserted 0.98
Building on this mechanism, we further propose ReKey, an algorithm that writes this signal by regression into a few key-value neurons of every block, and measures the site and strength of each edit on the model's own responses. method 0.83 0.89 0.05 asserted 1.00
Moreover, the base model fills every role of this recursion with forward passes alone, and its edits, written at different sites, compose in one checkpoint without refitting. method 0.78 0.74 0.08 asserted 1.00
Experimental results on ten visual reasoning benchmarks show that BaseRSI consistently improves a base vision-language model both in distribution and on held-out benchmarks. result 1.00 0.96 0.74 asserted 1.00
Our code is available at https://github.com/beita6969/BaseRSI. contribution 0.73 0.43 0.02 asserted 0.95

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99

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, 2026-11-19: 1 request, 6,442 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 (16)
Figure 1 of the paper
Figure 1. A harness gain that stays outside the base model is paid again on every request. BaseRSI lets the base model write what its harness learns into its own weights, round by round, and answer the original request alone. these gains still live in the request rather than in the model [60, 75], making them costly on every request and leaving the base model unable to build on what its harness learns. Along this path from the harness to the model, three main lines of work have emerged, as shown in
Figure 2 of the paper
Figure 2. Three lines of work and BaseRSI. (a) Harness evolution leaves harness-bound gains. (b) Model–harness co-evolution learns from black-box supervision. (c) Weight editing makes imprecise edits. (d) BaseRSI evolves the harness and internalizes each accepted gain into the base model from its own forward passes. evolution and harness distillation supervise the weights only at the output of the model, with trajectories, rewards, or out￾put distributions [38, 58, 75, 81]; what the harness changes inside the model stays hidden, so the signal travels back through every block by gradient…
Figure 3 of the paper
Figure 3. Overview of BaseRSI. Top: Recursive Harness Internalization runs an outer loop that evolves the harness and an inner loop in which the model with its harness teaches the model alone, round after round. Left: ReGap supervises every block with the teacher–student gap and keeps it as a low-dimensional update. Right: ReKey writes the gap into key–value neurons at a measured site and strength. objective there asks the edit to keep the base model as it is. Every edit is fitted on the base weights and takes its own neurons or the patch kernel, so the edits add up in one checkpoint without…
Figure 4 of the paper
Figure 4. Internalization, generalization, and process. (a) Gain over the base model with the harness in the request and with BaseRSI; ∆: their difference. (b) Gains of each edit alone (stacked) and of both; ∆: both minus the sum. (c) Restoration (share of the teacher divergence removed) on 28 held-out responses of the round-3 student, and test accuracy, from the base model to round 7. (d) Restoration by each round’s edit on responses of the teacher, the base model, and each round’s student; error bars: standard error. (e) Gain of BaseRSI per benchmark on four base models; bars: means. (f)…
Figure 5 of the paper
Figure 5. Fine-grained ablations, cost, and mechanisms of ReGap and ReKey (RQ3–RQ5). (a) Each row changes one choice of BaseRSI (setting in parentheses); Restor.: restoration; In sc.: gain (points) on the in-scope validation questions. (b) GPU hours and API tokens (filled: input, open: output). (c) Gain over the view edit per benchmark, distillation (x) against ReGap (y). (d) Divergence left after a step along the gap and after a unit step along other directions of equal norm; percentages: next-token agreement with the teacher. (e) Divergence removed by one block (stems) and by the first l…
Figure 6 of the paper
Figure 6. Rank, the bound of Proposition 3, and the steps in the thinking. (a) Restoration against the rank
Figure 7 of the paper
Figure 7. Cost of improving the base model against accuracy. (a) H800 GPU hours. (b) API tokens spent while improving. (c) ID and
Figure 8 of the paper
Figure 8. Counting the rings of a logo (VLMs Are Biased). The view edit alone answers four; the procedure edit turns it to five. Key observations. The question asks for the overlapping circles in the logo of a car, which VLMs Are Biased has re￾drawn with five rings (Figures 8 and 9). The base model rec￾ognizes an Audi and answers four in all ten runs; its think￾ing names the brand first and then counts “1, 2, 3, 4” along the rings it expects, so the count follows the prior rather than the image. The view edit raises the image tokens on the logo from 21 to 84, and the edited model still…
Figure 9 of the paper
Figure 9. Counting the rings of a logo, inside the model. Top: runs answered right, and runs whose thinking cites an instruction to trust what it sees, at each step of the procedure edit, 16 new runs each. Bottom: key value of the procedure edit (the gate) at every block along a response of BaseRSI: near zero over most of the question text, open through the thinking and the answer. Key observations. In 16 new runs per setting, none is right with the procedure edit off or at step 0.5, 1 at step 1.0, 14 at step 1.5, and all 16 at step 1.8, and 13 at step 2.0 (top). From none of 16 runs to all…
Figure 10 of the paper
Figure 10. Counting the stripes on a shoe (VLMs Are Biased). Each edit lifts the share of right runs, and together they reach 7/7. Key observations. The question asks for the white stripes in the logo on the left shoe, which VLMs Are Biased has drawn with four stripes where the brand has three (Fig￾ure 10). The base model is right in 2 of 10 runs and oth￾erwise answers three; its thinking reads “three white diag￾onal stripes” and settles on “three white bars”. Here both edits contribute. The view edit raises the image tokens on the shoe from 24 to 96 and the right answers to 3 of 5; in a…
Figure 11 of the paper
Figure 11. Counting the legs of a bird (VLMs Are Biased). Every closed-source run answers two; the view edit gives three in every run. Key observations. The question asks how many legs an an￾imal has, and the image shows a large flightless bird drawn with three legs (
Figure 12 of the paper
Figure 12. Counting the legs of a wolf (VLMs Are Biased). The base model sees “a normal walking dog”; the view edit counts five. Key observations. The image shows a wolf drawn with five legs (
Figure 13 of the paper
Figure 13. Counting black squares (BabyVision). At 72 image tokens a token is larger than a square; at 2,115 the model reads every row. Key observations. The question asks how many squares of a grid are black, 58 in all (
Figure 14 of the paper
Figure 14. Finding the odd one out (BabyVision). The base model gives ten positions in ten runs; the view edit finds the face. Key observations. The question asks for the row and the column of the one face that differs among 384 small faces, (7,8) in its answer format (
Figure 15 of the paper
Figure 15. Counting lights in a box (PerceptionBench, held out). The view edit lets the model locate the small lanterns by their coordinates. Key observations. The question asks how many lights lie in a given box of a game scene, six in all, on a benchmark that no part of BaseRSI saw (
Figure 16 of the paper
Figure 16. Reading a chart (ChartQAPro, held out). The base model and the other methods average the upper line; with the finer view the model averages the gap between the two lines, as the question asks, in all five runs. Key observations. The question asks for the average “un￾reasonable” wait time in a chart of the actual and the rea￾sonable wait for each year, that is, the mean gap between its two lines, 4.55 weeks (

Figures are the authors', reproduced from the paper.

References (84)
  1. [1] Lakshya A Agrawal, Shangyin Tan, Dilara Soylu, Noah Ziems, Rishi Khare, Krista Opsahl-Ong, Arnav Singhvi, Herumb Shandilya, Michael J Ryan, Meng Jiang, et al. Gepa: Reflective prompt evolution can outperform reinforcement learning. InInternational Conference on Learning Repre-... on Pith
  2. [2] James A Anderson. A simple neural network generating an interactive memory.Mathematical biosciences, 14(3-4): 197–220, 1972. 5 on Pith
  3. [3] Theodore Wilbur Anderson. Estimating linear restrictions on regression coefficients for multivariate normal distributions. The Annals of Mathematical Statistics, pages 327–351, 1951. 5 on Pith
  4. [4] Anthropic. Introducing Claude Haiku 5.5.https://www. anthropic.com/claude-haiku-5-5, 2026. 6 on Pith
  5. [5] Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. Re- fusal in language models is mediated by a single direction. Advances in Neural Information Processing Systems, 37: 136037–136083, 2024. 3 on Pith
  6. [6] Bhiman Kumar Baghel, Emma Jordan, Zheyuan Ryan Shi, and Xiang Lorraine Li. Resolving underedit & overedit with iterative & neighbor-assisted model editing.arXiv preprint arXiv:2503.11895, 2025. 2 on PitharXiv
  7. [7] David Bau, Steven Liu, Tongzhou Wang, Jun-Yan Zhu, and Antonio Torralba. Rewriting a deep generative model. In European conference on computer vision, pages 351–369. Springer, 2020. 5 on Pith
  8. [8] Ali Behrouz, Farnoosh Hashemi, Adel Javanmard, and Va- hab Mirrokni. Language models need sleep: Learning to self-modify and consolidate memories.arXiv preprint arXiv:2606.03979, 2026. 2 on PithPith's reviewarXiv
  9. [9] Lucas Beyer, Pavel Izmailov, Alexander Kolesnikov, Mathilde Caron, Simon Kornblith, Xiaohua Zhai, Matthias Minderer, Michael Tschannen, Ibrahim Alabdulmohsin, and Filip Pavetic. Flexivit: One model for all patch sizes. In 2023 IEEE/CVF Conference on Computer Vision and Pat- te... on Pith
  10. [10] Brian K Chen, Tianyang Hu, Hui Jin, Hwee Kuan Lee, and Kenji Kawaguchi. Exact conversion of in-context learning to model weights in linearized-attention transformers.arXiv preprint arXiv:2406.02847, 2024. 5 on PithPith's reviewarXiv
  11. [11] Liang Chen, Weichu Xie, Yiyan Liang, Hongfeng He, Hans Zhao, Zhibo Yang, Zhiqi Huang, Haoning Wu, Haoyu Lu, Yiping Bao, et al. Babyvision: Visual reasoning beyond lan- guage.arXiv preprint arXiv:2601.06521, 2026. 5 on PithPith's reviewarXiv
  12. [12] Roi Cohen, Eden Biran, Ori Yoran, Amir Globerson, and Mor Geva. Evaluating the ripple effects of knowledge edit- ing in language models.Transactions of the Association for Computational Linguistics, 12:283–298, 2024. 3 on Pith
  13. [13] Benoit Dherin, Michael Munn, Hanna Mazzawi, Michael Wunder, and Javier Gonzalvo. Learning without training: The implicit dynamics of in-context learning.arXiv preprint arXiv:2507.16003, 2025. 3, 4 on PithPith's reviewarXiv
  14. [14] Ali Ebrahimpour-Boroojeny, Xiaocong Yang, Samuel Schapiro, Zahra Borhani, Francisco R Ortega, Yujia Zheng, and ChengXiang Zhai. Resisting harmful instructions through training-free weight editing.arXiv preprint arXiv:2610.21189, 2026. 3 on PithPith's reviewarXiv
  15. [15] Carl Eckart and Gale Young. The approximation of one ma- trix by another of lower rank.Psychometrika, 1(3):211–218, on Pith
  16. [16] Jerome H Friedman. Greedy function approximation: a gra- dient boosting machine.Annals of statistics, pages 1189– 1232, 2001. 4, 5 on Pith
  17. [17] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. InEuropean Conference on Com- puter Vision, pages 148–166. Springer, 2024. 5 on Pith
  18. [18] Wenbo Gao, Zhaomou Song, Zhiyuan Ji, Renxi Liu, Xing Li, Xianzhi Yu, Xiaoguang Li, James Chung-wai Cheung, Weizhe Lin, and Yaoyuan Wang. Experience funnel: A state-policy alternating loop for self-evolving agents.arXiv preprint arXiv:2609.08919, 2026. 1, 2 on PithPith's reviewarXiv
  19. [19] Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. Transformer feed-forward layers are key-value memories. In Proceedings of the 2021 conference on empirical methods in natural language processing, pages 5484–5495, 2021. 5 on Pith
  20. [20] Adrian Goldwaser, Michael Munn, Javier Gonzalvo, and Benoit Dherin. Equivalence of context and parameter updates in modern transformer blocks.arXiv preprint arXiv:2511.17864, 2025. 3, 4 on PithPith's reviewarXiv
show 64 more
  1. [21] Jia-Chen Gu, Hao-Xiang Xu, Jun-Yu Ma, Pan Lu, Zhen-Hua Ling, Kai-Wei Chang, and Nanyun Peng. Model editing harms general abilities of large language models: Regular- ization to the rescue. InProceedings of the 2024 Confer- ence on Empirical Methods in Natural Language Processi... on Pith
  2. [22] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In2024 I... on Pith
  3. [23] Peter Hase, Mohit Bansal, Been Kim, and Asma Ghandehar- ioun. Does localization inform editing? surprising differ- ences in causality-based localization vs. knowledge editing in language models.Advances in Neural Information Pro- cessing Systems, 36:17643–17668, 2023. 2 on Pith
  4. [24] Prannay Hebbar, Yogendra Manawat, Samuel Verboomen, Alesia Ivanova, Selvam Palanimalai, Kunal Bhatia, and Vig- nesh Baskaran. Sia: Self improving ai with harness & weight updates.arXiv preprint arXiv:2605.27276, 2026. 2 on PithPith's reviewarXiv
  5. [25] Sture Holm. A simple sequentially rejective multiple test procedure.Scandinavian journal of statistics, pages 65–70, on Pith
  6. [26] Zeyu Huang, Yikang Shen, Xiaofeng Zhang, Jie Zhou, Wenge Rong, and Zhang Xiong. Transformer-patcher: One mistake worth one neuron.arXiv preprint arXiv:2301.09785, on PithPith's reviewarXiv
  7. [27] Alan Julian Izenman. Reduced-rank regression for the mul- tivariate linear model.Journal of multivariate analysis, 5(2): 248–264, 1975. 5 on Pith
  8. [28] Yubo Jiang, Fengying Xie, Zhiguo Jiang, and Haopeng Zhang. Distill skills into weights, not prompts: Abstract skills as privileged signals for on-policy self-distillation. arXiv preprint arXiv:2608.09826, 2026. 2 on PithPith's reviewarXiv
  9. [29] Teuvo Kohonen. Correlation matrix memories.IEEE trans- actions on computers, 100(4):353–359, 1972. 5 on Pith
  10. [30] Dharshan Kumaran, Demis Hassabis, and James L McClel- land. What learning systems do intelligent agents need? complementary learning systems theory updated.Trends in cognitive sciences, 20(7):512–534, 2016. 3 on Pith
  11. [31] Hyunin Lee, Jinglue Xu, Jeffrey Seely, Donghyun Lee, Matei Zaharia, and Yujin Tang. Recursive harness self- improvement.arXiv preprint arXiv:2607.15524, 2026. 1, 2, 3 on PithPith's reviewarXiv
  12. [32] Yoonho Lee, Roshen Nair, Qizheng Zhang, Kangwook Lee, Omar Khattab, and Chelsea Finn. Meta-harness: End- to-end optimization of model harnesses.arXiv preprint arXiv:2603.28052, 2026. 1, 2, 6 on PithPith's reviewarXiv
  13. [33] Changhao Li, Rushi Qiang, Jiawei Huang, Chenxiao Gao, Chao Zhang, Niao He, and Bo Dai. Revisiting dagger in the era of llm-agents.arXiv preprint arXiv:2605.12913, 2026. 4 on PithPith's reviewarXiv
  14. [34] Zichao Lin, Yifeng Xie, Bowen Qu, Haiming Wang, Jia Li, Haoning Wu, Yuhao Dong, Zuhao Yang, Jinguo Zhu, Haoyu Lu, et al. Perceptionbench: Evaluating atomic visual percep- tion in multimodal large language models.arXiv preprint arXiv:2607.24957, 2026. 5 on PithPith's reviewarXiv
  15. [35] Kaiyuan Liu, Zezhou Huang, Janvijay Singh, and Asli Ce- likyilmaz. What happens when the harness comes and goes? a controlled study of on-policy harness self-distillation. In NeurIPS 2026 Workshop on Transitioning from Pre-training to Post-training, 2026. 2 on Pith
  16. [36] David Lopez-Paz, L ´eon Bottou, Bernhard Sch ¨olkopf, and Vladimir Vapnik. Unifying distillation and privileged infor- mation.arXiv preprint arXiv:1511.03643, 2015. 3, 6 on PithPith's reviewarXiv
  17. [37] Bingjun Luo, Jialin Guo, and Siqi Li. Video-rsi: Recursive self-improvement of video understanding agents via harness evolution.arXiv preprint arXiv:2609.37950, 2026. 2 on PithPith's reviewarXiv
  18. [38] Haochen Luo, Yi Huang, Sichun Luo, Fengyuan Liu, Lei Li, Zefa Hu, Junlan Feng, and Qi Liu. Harness-aware self- evolving: Co-evolving model weights, harness, and task so- lutions.arXiv preprint arXiv:2607.03935, 2026. 1, 2 on PithPith's reviewarXiv
  19. [39] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hal- linan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: It- erative refinement with self-feedback.Advances in neural information processing systems, 36:46534–46594, 2023. 6 on Pith
  20. [40] Llew Mason, Jonathan Baxter, Peter Bartlett, and Marcus Frean. Boosting algorithms as gradient descent.Advances in neural information processing systems, 12, 1999. 4 on Pith
  21. [41] Ahmed Masry, Mohammed Saidul Islam, Mahir Ahmed, Aayush Bajaj, Firoz Kabir, Aaryaman Kartha, Md Tah- mid Rahman Laskar, Mizanur Rahman, Shadikur Rahman, Mehrad Shahmohammadi, et al. Chartqapro: A more diverse and challenging benchmark for chart question answering. In Findings ... on Pith
  22. [42] Hanna Mazzawi, Benoit Dherin, Michael Munn, Adrian Goldwaser, Michael Wunder, and Javier Gonzalvo. Transmuting prompts into weights.arXiv preprint arXiv:2510.08734, 2025. 3, 4 on PithPith's reviewarXiv
  23. [43] James L McClelland, Bruce L McNaughton, and Randall C O’Reilly. Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory. Psychological review, 102(3):419, 1995. 3 on Pith
  24. [44] Kevin Meng, David Bau, Alex Andonian, and Yonatan Be- linkov. Locating and editing factual associations in gpt.Ad- vances in neural information processing systems, 35:17359– 17372, 2022. 1, 2, 5 on Pith
  25. [45] Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. Mass-editing memory in a trans- former.arXiv preprint arXiv:2210.07229, 2022. 1, 2 on PithPith's reviewarXiv
  26. [46] Michael Merzin, Yehonatan Elisha, Oren Barkan, Yonatan Toib, and Noam Koenigstein. In-context compilation: En- coding context-induced behavioral shift into model weights. InFindings of the Association for Computational Linguis- tics: EMNLP 2026, pages 20777–20796, 2026. 1, 3 on Pith
  27. [47] OpenAI. Introducing GPT-6 Sol and Luna.https:// openai.com/index/introducing-gpt-6-sol- and-luna/, 2026. 6 on Pith
  28. [48] Haowen Pan, Xiaozhi Wang, Yixin Cao, Zenglin Shi, Xun 10 Yang, Juanzi Li, and Meng Wang. Precise localization of memories: A fine-grained neuron-level knowledge editing technique for llms. InInternational Conference on Learning Representations, pages 482–502, 2025. 2 on Pith
  29. [49] Judea Pearl. Direct and indirect effects. InProbabilistic and causal inference: the works of Judea Pearl, pages 373–392 on Pith
  30. [50] PixelProof. PixelProof: Visual question generation through forward–inverse agreement.https://github.com/ ProjectPixelProof/PixelProof, 2026. 6 on Pith
  31. [51] Tao Ren, Weiyao Luo, Hui Yang, Rongzhi Zhu, Xiang Huang, Yuchuan Wu, Bingxue Chou, Jieping Ye, Jiafeng Liang, Yongbin Li, et al. Scaling self-evolving agents via parametric memory.arXiv preprint arXiv:2606.04536, 2026. 2 on PithPith's reviewarXiv
  32. [52] St ´ephane Ross and Drew Bagnell. Efficient reductions for imitation learning. InProceedings of the thirteenth inter- national conference on artificial intelligence and statistics, pages 661–668. JMLR Workshop and Conference Proceed- ings, 2010. 4 on Pith
  33. [53] St ´ephane Ross, Geoffrey Gordon, and Drew Bagnell. A re- duction of imitation learning and structured prediction to no- regret online learning. InProceedings of the fourteenth inter- national conference on artificial intelligence and statistics, pages 627–635. JMLR Workshop a... on Pith
  34. [54] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of math- ematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024. 6 on PithPith's reviewarXiv
  35. [55] Noam Shazeer. Glu variants improve transformer.arXiv preprint arXiv:2002.05202, 2020. 3 on PithPith's reviewarXiv
  36. [56] Haozhan Shen, Kangjia Zhao, Tiancheng Zhao, Ruochen Xu, Zilun Zhang, Mingwei Zhu, and Jianwei Yin. Zoom- eye: Enhancing multimodal llms with human-like zooming capabilities through tree-based image exploration. InPro- ceedings of the 2025 Conference on Empirical Methods in Nat... on Pith
  37. [57] Yueqi Song, Tianyue Ou, Yibo Kong, Zecheng Li, Graham Neubig, and Xiang Yue. Visualpuzzles: Decoupling multi- modal reasoning evaluation from domain knowledge.arXiv preprint arXiv:2504.10342, 2025. 5 on PithPith's reviewarXiv
  38. [58] Jiaqi Su, Guibin Zhang, Jie-Jing Shao, Cong Pang, Tiankuo Yao, Ivor Tsang, Haiyan Yin, and Shuicheng Yan. Coevo-rsi: When harnesses evolve, which models should follow?arXiv preprint arXiv:2610.25078, 2026. 2 on PithPith's reviewarXiv
  39. [59] Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khat- tab. Sparks of weight-space programming: Llms can natively read & write neural network weights.MIT CSAIL research blog. Published November, 2:2026, 2026. 3 on Pith
  40. [60] Zihan Tan, Leixin Sun, Zitong Shi, Yitao Liu, Jiajun Wu, Nathaniel Brooks, Jiaru Qian, Xiaoran Shang, Suyuan Huang, Yi Ding, et al. Metarsi/rsi2: A meta-recursive self-improving system for recursive self-improving systems themselves.arXiv preprint arXiv:2609.06396, 2026. 1, 2 on PithPith's reviewarXiv
  41. [61] Gemma Team, Sherif El Abd, Vaibhav Aggarwal, Robin Al- gayres, Alek Andreev, Olivier Bachem, Ian Ballantyne, Cor- mac Brick, Victor C˘arbune, Michelle Casbon, et al. Gemma 4 technical report.arXiv preprint arXiv:2607.02770, 2026. 6 on PithPith's reviewarXiv
  42. [62] Vladimir Vapnik and Akshay Vashist. A new learning paradigm: Learning using privileged information.Neural networks, 22(5-6):544–557, 2009. 3, 6 on Pith
  43. [63] An V o, Khai-Nguyen Nguyen, Mohammad Reza Tae- siri, Vy Tuong Dang, Anh Totti Nguyen, and Daeyoung Kim. Vision language models are biased.arXiv preprint arXiv:2505.23941, 2025. 5 on PithPith's reviewarXiv
  44. [64] Siting Wang, Minnan Pei, Luoyang Sun, Cheng Deng, Yuchen Li, Kun Shao, Zheng Tian, Haifeng Zhang, and Jun Wang. Spatialviz-bench: A cognitively-grounded bench- mark for diagnosing spatial visualization in mllms. InIn- ternational Conference on Learning Representations, pages 4... on Pith
  45. [65] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reason- ing in language models.arXiv preprint arXiv:2203.11171, on PithPith's reviewarXiv
  46. [66] Zirui Wang, Mengzhou Xia, Luxi He, Howard Chen, Yitao Liu, Richard Zhu, Kaiqu Liang, Xindi Wu, Haotian Liu, Sad- hika Malladi, et al. Charxiv: Charting gaps in realistic chart understanding in multimodal llms.Advances in Neural In- formation Processing Systems, 37:113569–11369... on Pith
  47. [67] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large lan- guage models.Advances in neural information processing systems, 35:24824–24837, 2022. 6 on Pith
  48. [68] Peng Xia, Rujun Han, Zifeng Wang, Yanfei Chen, Yufan Zhuang, Yoonho Lee, Chengsong Huang, Han Yu, Zhongy- ing CuiZhu, Yifei Ming, et al. Rrsi: Regularized recur- sive self-improvement of agent harnesses.arXiv preprint arXiv:2609.24972, 2026. 1, 2, 6 on PithPith's reviewarXiv
  49. [69] Chejian Xu, Shujian Zhang, Wenxuan Zhou, John Lambert, Beck Chen, Bo Li, Andrew Hard, and Rajiv Mathews. Evo2: Evolving the evolver through weight-harness co-evolution. arXiv preprint arXiv:2610.13025, 2026. 2 on PithPith's reviewarXiv
  50. [70] Guoyang Xu and Hao Chen. Videoharness-rsi: Recur- sive harness self-improvement for long-video understand- ing with frozen vision-language models.arXiv preprint arXiv:2608.24302, 2026. 1, 2, 6 on PithPith's reviewarXiv
  51. [71] Weiye Xu, Jiahao Wang, Weiyun Wang, Zhe Chen, Wengang Zhou, Aijun Yang, Lewei Lu, Houqiang Li, Xiaohua Wang, 11 Xizhou Zhu, et al. Visulogic: A benchmark for evaluating vi- sual reasoning in multi-modal large language models. InIn- ternational Conference on Learning Representa... on Pith
  52. [72] Bangji Yang, Jiajun Fan, Zhaoyu Zhu, Hongbo Ma, and Ge Liu. Training-free on-policy distillation.arXiv preprint arXiv:2610.22236, 2026. 1, 3 on PithPith's reviewarXiv
  53. [73] Yifan Yang, Ziyang Gong, Weiquan Huang, Qihao Yang, Zi- wei Zhou, Zisu Huang, Yan Li, Xuemei Gao, Qi Dai, Bei Liu, et al. Skillopt: Executive strategy for self-evolving agent skills.arXiv preprint arXiv:2605.23904, 2026. 6 on PithPith's reviewarXiv
  54. [74] Ruilin Yao, Xinyu Qiu, Tianyu Zou, Yi Rong, Shengwu Xiong, Shuheng Shen, and Changhua Meng. Learning to look beyond screens: Transferable visual observation from gui grounding.arXiv preprint arXiv:2610.37093, 2026. 6 on PithPith's reviewarXiv
  55. [75] Haoran Ye, Yuxing Lu, Haonan Dong, Zhaochen Su, and Guojie Song. Harness-zero: Harness distillation via agent- as-harness.arXiv preprint arXiv:2609.24974, 2026. 1, 2 on PithPith's reviewarXiv
  56. [76] Zhou Yu, Bin Bi, Shiva Kumar Pentyala, Shubham Mehro- tra, Sougata Chaudhuri, Shilpa Bhagavath, Zeyuan Chen, Ran Xu, Phil Mui, James Zhu, et al. Co-evolving harnesses and models: On-policy correction helps weaker models catch up where imitation fails.arXiv preprint arXiv:2609.09134, on PithPith's reviewarXiv
  57. [77] Yuchen Zeng, Giorgio Severi, Blake Bullwinkel, Eric Horvitz, and Dimitris Papailiopoulos. Lazare, veni foras: Reprogramming dormant neurons to backdoor ai agents. arXiv preprint arXiv:2611.15118, 2026. 3 on PithPith's reviewarXiv
  58. [78] Mengqi Zhang, Xiaotian Ye, Qiang Liu, Pengjie Ren, Zhumin Chen, et al. Uncovering overfitting in large lan- guage model editing. InInternational Conference on Learn- ing Representations, pages 79550–79568, 2025. 3 on Pith
  59. [79] Qizheng Zhang, Changran Hu, Shubhangi Upasani, Boyuan Ma, Fenglu Hong, Vamsidhar Kamanuru, Jay Rainton, Chen Wu, Mengmeng Ji, Hanchen Li, et al. Agentic context en- gineering: Evolving contexts for self-improving language models. InInternational Conference on Learning Represen... on Pith
  60. [80] Fanyu Zhao, Ruike Cao, Liang Dong, Fugen Yao, Jian Xu, Guanjun Jiang, Han Zhang, Yifei Zhao, and Yin- sheng Li. Rpmem: Learning long-term recurrent paramet- ric memory across sessions for llm agents.arXiv preprint arXiv:2609.23466, 2026. 2 on PithPith's reviewarXiv
  61. [81] Zhengyang Zhao, Lu Ma, and Wentao Zhang. Training with harnesses: On-policy harness self-distillation for complex reasoning.arXiv preprint arXiv:2605.08741, 2026. 2 on PithPith's reviewarXiv
  62. [82] Tailin Zhou. Hierarchical self-improvement: A framework for task-specific evolvable agent harnesses.arXiv preprint arXiv:2608.08466, 2026. 1, 2 on PithPith's reviewarXiv
  63. [83] Adam Zweiger, Jyo Pari, Han Guo, Yoon Kim, and Pulkit Agrawal. Self-adapting language models.Advances in Neu- ral Information Processing Systems, 38:74084–74115, 2026. 2 12 BaseRSI: Recursive Self-Improvement of a Base Vision-Language Model via Training-Free Harness Internaliz... on Pith
  64. [84] [. . . ] The Audi emblem is four rings. Yes.”BaseRSI, run 0: “Looking at the image, the logo shows overlapping rings. Let me count them individually [. . . ] Counting: leftmost ring, then sec- ond, third, fourth, fifth. [. . . ] Let me double check by mentally tracing: circle ... on Pith

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