Pith. sign in

REVIEW 5 major objections 5 minor 72 references

Large pre-trained models can improve their own zero-shot accuracy by 1–3 points through inference-time edits to residual and cross-head shortcuts, requiring no data, no gradients, and no weight updates.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 05:18 UTC pith:BXVQBDF6

load-bearing objection Head-mixing shortcut is a real new idea, but the VLM headline gains are partly a test-set selection artifact and the α definition contradicts itself — worth a serious referee, not acceptance as is. the 5 major comments →

arxiv 2607.13395 v1 pith:BXVQBDF6 submitted 2026-07-15 cs.LG

Self-Improving is Often Sudden: Enlightenment-style Finetuning for Large-Scale Models

Jing-Xiao Liao , Tianwei Zhang , Yu-Hao Jiang , Feifei Zhang , Hang-Cheng Dong , Feng-Lei Fan This is my paper
classification cs.LG
keywords training-free fine-tuningself-improvementattention head mixingresidual dampingzero-shot accuracyvision-language modelsquantization recoveryattention sink
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to show that a frozen, fully trained large model can 'enlighten' itself at inference time—suddenly getting better on benchmarks with no extra data, no gradient computation, and no weight updates. The proposed mechanism is architectural: instead of modifying weights, it rescales existing residual shortcuts and, in language models, injects new cross-head connections that mix attention-head outputs. Across three large language models and three vision-language models, the paper reports average zero-shot gains of about 1.3 to 1.9 points on language tasks and 1.9 to 3.1 points on multimodal tasks, and it partially restores accuracy lost to aggressive quantization. A sympathetic reader would care because the method points to a third lever—network structure—alongside weights and data for improving large models at deployment time.

Core claim

On the paper's own terms, the discovery is that the representational capacity of a pre-trained transformer is not fully fixed by its weights: a handful of scalar coefficients placed on shortcut paths can re-allocate that capacity. For language models, the paper scales each attention head's attention to the initial token using a per-head factor drawn from a Chebyshev polynomial, then adds the first head's output—scaled by a mixing rate—into every other head, creating shortcuts that did not exist during training. For vision-language models, it multiplies every residual connection in the decoder, both attention and feed-forward sub-blocks, by a single scalar, damping over-active sublayer output

What carries the argument

The central mechanism is shortcut modulation: treating residual connections and cross-head paths as adjustable gates rather than fixed wires. For LLMs, the key object is the 'head-mixing shortcut': the output of attention head 0 is scaled and added into every other head's output, while each head's attention to the first token is pre-scaled by a per-head factor fitted through a third-order Chebyshev polynomial; this recalibrates the well-known attention-sink behavior at per-head granularity. For VLMs, the key object is a single scalar that multiplies all residual connections in the decoder's attention and feed-forward blocks, uniformly damping sublayer outputs while preserving the residual st

Load-bearing premise

The load-bearing premise is that the scalar hyperparameters were chosen without peeking at the same test benchmarks used to report the gains; if those values were fitted to the benchmark results, the central claim of training-free, data-free self-improvement collapses into benchmark fitting.

What would settle it

Take any reported benchmark, split it into a tuning set and a held-out set, select the scalar hyperparameters only on the tuning set, and evaluate on the held-out set with the same frozen model; if the gains disappear or reverse on held-out data, the 'latent capacity' explanation is not supported.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If valid, training-free post-tuning can be viewed as a structural operation, not a weight operation: any frozen transformer may carry untapped representational capacity recoverable by shortcut edits.
  • For LLMs, the approach gives per-head control over attention-sink distribution, yielding consistent average gains of about 1.3 to 1.9 points across three model families, in contrast to global suppression of the initial token.
  • For VLMs, a single residual-damping scalar improves average zero-shot accuracy by about 1.9 to 3.1 points across five multimodal benchmarks, with the largest gains on perception-heavy categories.
  • For aggressively quantized models, the same shortcut edits recover part of the accuracy lost to low-bit compression, including a large recovery on one question-answering benchmark, suggesting a cheap complement to post-quantization recovery.
  • The two instantiations are architecture-specific: head-mixing hurts VLMs and residual damping hurts LLMs, so the mechanism must be matched to the model family.
  • The method introduces no new parameters but relies on a small set of scalar hyperparameters that must be set per model family.
  • The 'sudden enlightenment' framing implies that a model's latent capability is a structural property, not just a parametric one, which could reshape how practitioners think about inference-time adaptation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the reported gains are small relative to typical benchmark variance and depend on a few scalars, the most direct extension is a stress test: choose the scalars on a development set and report held-out gains; the paper's own parameter sweeps are not a substitute for this protocol.
  • The paper's activation analysis suggests over-activation concentrates in late decoder layers; a natural extension is layer-dependent damping, which the paper's own layer-range experiments hint could specialize gains per task.
  • The cross-architecture asymmetry implies the two interventions correct different failure modes—attention-sink dominance in LLMs versus late-layer visual noise in VLMs—so a gated or input-dependent combination might outperform either single scalar on mixed reasoning-perception tasks.
  • For quantization, the method's ability to recover accuracy suggests it could serve as a cheap, inference-only complement to quantization-aware training; whether that holds under held-out calibration remains untested.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes "Enlightenment," a training-free post-tuning paradigm that modifies existing or inserts new shortcut connections rather than updating weights. For LLMs it scales the BOS attention of each head via a Chebyshev-fitted adaptive factor and injects head-0 output into other heads (head mixing). For VLMs it multiplies every decoder residual sublayer output by a single scalar eta, i.e., residual damping. The authors report average zero-shot accuracy gains of about +1.3 to +1.9 points on three LLMs, +2.2 to +3.1 points on three VLMs, and a partial recovery of W4A4C2-quantized accuracy, claiming the procedure requires no data, no gradients, and no weight updates.

Significance. If the central claim is established, the contribution is interesting: it identifies architectural shortcut modification as an axis for inference-time adaptation that is orthogonal to the attention-weight-modification methods that dominate training-free literature. The paper's breadth is a strength: three LLMs, three VLMs, a quantized model, multiple benchmarks, per-category analyses, and open-sourced code. The high-level idea is simple enough to be practical and falsifiable. However, the current evidence does not establish the central "data-free self-improvement" claim, because the most important scalar hyperparameters appear to be selected on the same benchmarks that later appear as headline results. The paper is therefore better read as a promising mechanism with a protocol problem than as a validated method.

major comments (5)
  1. [Sec. 5.1, Fig. 6, Table 11] The VLM headline result is selected on the test set. Figure 6 sweeps eta on MMBench-300 and shows Qwen2-VL-7B peaking at eta=0.5 with about a 6% gain; Sec. 5.1 then fixes eta=0.5 for all VLM runs; Table 11 reports MMBench-300 (79.33 to 85.33 for Qwen2-VL-7B) as a headline result. Thus the largest reported gain is on the same subset used for hyperparameter selection. The paper does not provide a held-out protocol or a validation split. Because the central claim is that the method is data-free and training-free, not benchmark-fitted, this is a load-bearing issue. The authors must re-run with eta chosen on a disjoint validation set or report the full eta sweep on all benchmarks.
  2. [Sec. 3.1, Eq. (7), Table 8] Eq. (7) states that the combined head output is [\tilde A0 V0, alpha \tilde A0 V0 + \tilde A1 V1, ...]. Under this equation alpha=0 gives the original head outputs, while alpha=1 injects the full head-0 output into every other head. The text following Eq. (7) says "When alpha=1, the attention module remains unmodified," which contradicts the displayed equation, and Table 8 labels alpha=1 as "disabled." This is not a typo in a peripheral equation: the LLM recipe, the ablation, and the interpretation of "disabled" all depend on which definition is used. The equation and the text/table must be reconciled.
  3. [Sec. 4.1 vs. Sec. 3.1] The method section defines per-head adaptive scaling factors from a Chebyshev fit over initial-head attention values, but Sec. 4.1 says the main LLM results use "attention head mixing with uniform per-head rates." These are different procedures. If Table 4 uses uniform rates, the adaptive mechanism (the methodological novelty of the LLM instantiation) is not actually evaluated in the main results; if Table 4 uses the adaptive rates, the experimental description is wrong. The paper must specify exactly which gamma configuration was used in each table and report the per-head rates actually applied.
  4. [Sec. 4.1, Table 5] The quantized-model result uses beta=0.5, but Eq. (14) describes beta as a task-dependent factor. No held-out selection protocol for beta is reported, and Table 5 is then presented as a main demonstration that Enlightenment recovers quantization loss. The same benchmark-fitting concern that applies to eta applies here, on a smaller scale but with an equally undisclosed selection procedure. The authors should provide a sensitivity analysis over beta for all tasks in Table 5 or specify a predetermined default.
  5. [Sec. 4.3, Tables 8 and 10] The hyperparameter-insensitivity claim is supported only by experiments on 100 BoolQ samples with a single seed, no confidence interval, and no explicit baseline column. This is too weak to establish that the defaults are safe across unseen benchmarks, and it does not resolve the concern that the reported gains may be concentrated at specially chosen hyperparameters. A small multi-benchmark sensitivity analysis with variance estimates would be needed to support the claim that the method is "insensitive to most hyperparameters."
minor comments (5)
  1. [Notation] Eq. (15) and Sec. 5.3 use eta for the VLM scalar, but Sec. 5.1 and Table 13 use gamma=0.5. Please unify the notation.
  2. [Sec. 4.1] The sentence "with all decoder layers applied scalar-modulated factor for Enlightenment" appears in the LLM setup, but the LLM instantiation uses head mixing. This appears to be a copy-paste error from the VLM section.
  3. [Table 11 and Sec. 5.2] The table reports Avg. Delta vs base as +2.16 for Qwen2-VL-7B, but the text says +1.92. These numbers should be reconciled.
  4. [Table 12 vs. Table 11] Table 12 analyzes a 1,292-sample MMBench set while Table 11 reports MMBench-300. The relationship between these subsets should be stated explicitly, and the use of a 300-sample subset for the headline table should be justified.
  5. [Typos] Table 8 header reads "Ploynomial degree"; Figure 8 caption reads "actication reduction"; the abstract has a line-break artifact in "Enlighten- ment." Please proofread.

Circularity Check

1 steps flagged

VLM headline MMBench gain is a fit: η=0.5 is chosen by sweeping MMBench-300, then the same MMBench-300 subset is reported as the headline Table 11 result.

specific steps
  1. fitted input called prediction [Section 5.3 / Figure 6; Section 5.1; Section 5.2 / Table 11]
    "Sweep η on MMBench-300 across 4 primary models. η=1.0 is the baseline. ... Qwen2-VL-7B is the sole exception, displaying a clear inverted-U pattern that peaks at η=0.5 with a 6% performance gain. ... For Enlightenment, we apply a single shared scalar modulator γ=0.5 to all residual connections ... Table 11 ... MMBench 300 ... 79.33 -> 85.33."

    The single VLM hyperparameter η is selected by maximizing accuracy on the 300-sample MMBench subset: Figure 6 shows that Qwen2-VL-7B peaks at η=0.5 with a 6% gain. Table 11 then reports the same MMBench-300 subset as a headline result, with the same Qwen2-VL-7B entry going from 79.33 to 85.33. That reported gain is the value of the selection objective on the fitting set, not an independent prediction from a fixed, zero-tuning intervention. The MMBench evidence for the central 'training-free, data-free self-improvement' claim is therefore fit-then-report on the same benchmark, even though other benchmarks are not explicitly shown to be selection targets.

full rationale

The method itself is an architectural intervention (head-mixing for LLMs, scalar residual scaling for VLMs) and most of the reported evaluations are not theoretically derived from the method's own assumptions, so the core derivation chain is not circular in the usual Eq.-X-equals-Eq.-Y sense. The load-bearing circularity is narrower: the VLM scalar η=0.5 is chosen by sweeping MMBench-300, and the paper then reports MMBench-300 as a headline success. That specific result is a fitted input called a prediction; it does not independently validate the zero-tuning claim. I score 6, not higher, because the LLM experiments and the non-MMBench VLM benchmarks are not shown to be hyperparameter selection targets, and the modifications are not defined by the benchmark outcomes. I do not count the LLM β=0.5 setting for the quantized run as a separate circular step: β is called 'task-dependent' (Eq. 14), but the paper does not disclose a sweep, so treating this as fitted-to-test would be speculation. I also flag for correctness/reproducibility, not circularity, that the text 'When α=1, the attention module remains unmodified' contradicts Eq. 7, where α=0 gives the original head outputs and α=1 injects the full head-0 output into every other head. The self-citations in the paper (e.g., [14] for shortcut efficacy) are not load-bearing; the main results are benchmark evaluations rather than a chain of authority, so this does not raise the score further.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim rests on several hand-chosen scalars (η, α, β, Chebyshev anchors) that are tuned on the same benchmarks later reported, plus two empirical hypotheses (attention sink, over-active sublayers) and one unproven architectural assumption (head-0 output helps all heads). No new physical or computational entity requiring independent falsification is introduced.

free parameters (5)
  • η (VLM residual scalar) = 0.5 (default)
    Selected from MMBench-300 sweep (Figure 6); inverted-U for Qwen2-VL peaks at 0.5. Used in all VLM tables; no held-out protocol.
  • α (LLM head-mixing rate) = 0.1 (default)
    Grid-searched on BoolQ (Table 8); sensitivity is flat except α=1 called 'disabled'. Eq. 7 makes α=0 the unmodified case, so the implemented semantics are ambiguous.
  • β (task-dependent Chebyshev factor) = 1 (default); 0.5 for W4A4C2 run
    Introduced in Eq. 14 to handle domain/task shift; set per model/quantization setting, so it is a tuned constant rather than a derived value.
  • Chebyshev anchor points = {(0.05,1.5),(0.15,1.1),(0.30,1),(0.50,0.7)}
    Hand-chosen in Eq. 11 to force attenuation of high BOS attention; these anchors determine every γ_h value.
  • Polynomial degree d = 3
    Default degree; d=1 yields 92.0 vs d=3 94.0 on BoolQ-100, so the default is selected by performance on the evaluation set.
axioms (4)
  • domain assumption Attention-sink hypothesis: the BOS token receives excessive attention and dampening it improves task accuracy.
    Used to justify Eqs. 4–5 and head-mixing; cited from Refs. [3,17,19,26,58] but not re-established here. Table 9 shows scaling alone is unstable.
  • domain assumption Over-active sublayer hypothesis: deep decoder sublayer outputs are noisy and can be safely uniformly damped.
    Used for Eq. 15; only supported by a one-model activation analysis (Figures 8–9), not a general law.
  • ad hoc to paper Head 0 is a useful shared feature for all other attention heads.
    Eq. 7 injects head0 output into every head; the ablation in Table 9 shows HeadMix alone regresses to 80.3% from 80.4%, so this premise is not independently supported.
  • domain assumption Perplexity-based label-token scoring is a faithful measure of task ability.
    All LLM classification results use log-likelihood over templates rather than open-ended generation; this may not reflect actual self-improvement in deployment.

pith-pipeline@v1.3.0-alltime-deepseek · 19787 in / 17614 out tokens · 189657 ms · 2026-08-02T05:18:34.580162+00:00 · methodology

0 comments
read the original abstract

The pursuit of autonomously self-improving models has attracted growing interest in the era of large-scale foundation models. Drawing inspiration from the concept of "enlightenment" or "aha moment" in human brain, we hypothesize that large models exhibit an analogous enlightenment phenomenon-a latent capacity for sudden capability boost. Then, we propose Enlightenment, a novel training-free post-tuning paradigm for large-scale models. Our approach modifies shortcuts for key modules/layers without weight updates, while existing training-free ones predominantly manipulate attention weights. We introduce two architecture-specific instantiations: i) For large language models, we propose attention head-mixing shortcuts that recalibrate attention weights by linking the initial attention head's output to all other target heads, modulated by an adaptive scaling factor initialization strategy. ii) For vision-language models, we apply a lightweight scalar-modulated factor to residual connections in the decoder layers, regulating information flow. Extensive experiments show that Enlightenment efficiently unlocks the latent potential of pre-trained networks, yielding remarkable performance improvements across diverse benchmarks and models.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

72 extracted references · 2 linked inside Pith

  1. [1]

    Perspectives on Science 14(2):153–188

    Arikha N (2006) Form and function in the early enlight- enment. Perspectives on Science 14(2):153–188

    2006

  2. [2]

    arXiv preprint arXiv:251121631

    Bai S, Cai Y , Chen R, Chen K, Chen X, Cheng Z, Deng L, Ding W, Gao C, Ge C, et al (2025) Qwen3-vl techni- cal report. arXiv preprint arXiv:251121631

    2025

  3. [3]

    Barbero F, Arroyo A, Gu X, Perivolaropoulos C, Bronstein M, Veli ˇckovi´c P, Pascanu R (2025) Why do llms attend to the first token? arXiv preprint arXiv:250402732

    2025

  4. [4]

    In: Classification, clustering, and data analysis: recent advances and applications, Springer, pp 281–288

    Billard L, Diday E (2002) Symbolic regression analysis. In: Classification, clustering, and data analysis: recent advances and applications, Springer, pp 281–288

    2002

  5. [5]

    In: Proceedings of the AAAI conference on artificial in- telligence, vol 34, pp 7432–7439

    Bisk Y , Zellers R, Gao J, Choi Y (2020) Piqa: Reason- ing about physical commonsense in natural language. In: Proceedings of the AAAI conference on artificial in- telligence, vol 34, pp 7432–7439

    2020

  6. [6]

    Advances in neural information processing systems 37:57817–57840

    Chen C, Zhu J, Luo X, Shen HT, Song J, Gao L (2024) Coin: A benchmark of continual instruction tuning for multimodel large language models. Advances in neural information processing systems 37:57817–57840

    2024

  7. [7]

    arXiv preprint arXiv:260312658

    Chen H, Sun Z, Ye H, Li K, Lin X (2026) Con- tinual learning in large language models: Meth- ods, challenges, and opportunities. arXiv preprint arXiv:260312658

    2026

  8. [8]

    Proceedings of the National Academy of Sciences 114(18):4637–4642

    Chen WJ, Krajbich I (2017) Computational model- ing of epiphany learning. Proceedings of the National Academy of Sciences 114(18):4637–4642

    2017

  9. [9]

    In: International Conference on Ma- chine Learning, PMLR, pp 6621–6642

    Chen Z, Deng Y , Yuan H, Ji K, Gu Q (2024) Self-play fine-tuning converts weak language models to strong language models. In: International Conference on Ma- chine Learning, PMLR, pp 6621–6642

    2024

  10. [10]

    Clark C, Lee K, Chang MW, Kwiatkowski T, Collins M, Toutanova K (2019) BoolQ: Exploring the surprising difficulty of natural yes/no questions. In: Proceedings of the 2019 Conference of the North American Chap- ter of the Association for Computational Linguistics: Human Language Technologies, V olume 1 (Long and Short Papers), Association for Computational...

    doi:10.18653/v1/n19-1300 2019

  11. [11]

    arXiv preprint arXiv:180305457

    Clark P, Cowhey I, Etzioni O, Khot T, Sabharwal A, Schoenick C, Tafjord O (2018) Think you have solved question answering? try arc, the ai2 reasoning chal- lenge. arXiv preprint arXiv:180305457

    2018

  12. [12]

    URLhttps://arxiv.org/abs/2501.12948, 2501.12948

    DeepSeek-AI (2025) Deepseek-r1: Incentivizing rea- soning capability in llms via reinforcement learning. URLhttps://arxiv.org/abs/2501.12948, 2501.12948

    Pith/arXiv arXiv 2025

  13. [13]

    Med-X 3(1):2

    Fan FL, Li Y , Zeng T, Wang F, Peng H (2025) To- wards neuroai: introducing neuronal diversity into ar- tificial neural networks. Med-X 3(1):2

    2025

  14. [14]

    IEEE Transactions on Ar- tificial Intelligence 7(3):1223–1237, DOI 10.1109/TAI

    Fan FL, Li ZY , Xiong H, Zeng T (2026) On expressivity of height in neural networks. IEEE Transactions on Ar- tificial Intelligence 7(3):1223–1237, DOI 10.1109/TAI. 2025.3592162

    arXiv 2026

  15. [15]

    Advances in Neural Information Pro- cessing Systems 38

    Fu C, Chen P, Shen Y , Qin Y , Zhang M, Lin X, Yang J, Zheng X, Li K, Sun X, et al (2026) Mme: A com- prehensive evaluation benchmark for multimodal large language models. Advances in Neural Information Pro- cessing Systems 38

    2026

  16. [16]

    International Journal of Computer Vision 131(7):1857–1874

    Gou J, Xiong X, Yu B, Du L, Zhan Y , Tao D (2023) Multi-target knowledge distillation via student self- reflection. International Journal of Computer Vision 131(7):1857–1874

    2023

  17. [17]

    In: International Confer- ence on Learning Representations, vol 2025, pp 97114– 97144

    Gu X, Pang T, Du C, Liu Q, Zhang F, Du C, Wang Y , Lin M (2025) When attention sink emerges in language models: An empirical view. In: International Confer- ence on Learning Representations, vol 2025, pp 97114– 97144

    2025

  18. [18]

    In: Proceedings of the 58th annual meeting of the asso- ciation for computational linguistics, pp 8342–8360

    Gururangan S, Marasovi ´c A, Swayamdipta S, Lo K, Beltagy I, Downey D, Smith NA (2020) Don’t stop pre- training: Adapt language models to domains and tasks. In: Proceedings of the 58th annual meeting of the asso- ciation for computational linguistics, pp 8342–8360

    2020

  19. [19]

    In: The Fourteenth International Conference on Learn- ing Representations, URLhttps://openreview

    Han F, Yu X, Tang J, Rao D, Du W, Ungar L (2026) Zerotuning: Unlocking the initial token’s power to enhance large language models without training. In: The Fourteenth International Conference on Learn- ing Representations, URLhttps://openreview. net/forum?id=EdkQ14FiiO

    2026

  20. [20]

    Advances in neural information pro- cessing systems 36:45870–45894

    Hao S, Liu T, Wang Z, Hu Z (2023) Toolkengpt: Aug- menting frozen language models with massive tools via tool embeddings. Advances in neural information pro- cessing systems 36:45870–45894

    2023

  21. [21]

    IEEE Transactions on Image Processing

    He J, Guo H, Zhu K, Tang M, Wang J (2026) Continual instruction tuning for large multimodal models. IEEE Transactions on Image Processing

    2026

  22. [22]

    In: Proceedings of the 2023 conference on empirical meth- ods in natural language processing, pp 1051–1068

    Huang J, Gu S, Hou L, Wu Y , Wang X, Yu H, Han J (2023) Large language models can self-improve. In: Proceedings of the 2023 conference on empirical meth- ods in natural language processing, pp 1051–1068

    2023

  23. [23]

    Islam A, Bouzerdoum A, Belhaouari SB (2026) Bio- inspired adaptive neurons for dynamic weighting in ar- tificial neural networks. AI Open

    2026

  24. [24]

    arXiv preprint Enlightenment-style Tuning for Large Models 15 arXiv:260630168

    Jiang K, Zhu R, Huang S, Zhang H, Li X (2026) Latent noise mask for reducing visual redundancy in multimodal large language models. arXiv preprint Enlightenment-style Tuning for Large Models 15 arXiv:260630168

    2026

  25. [25]

    International Journal of Computer Vision 134(4):190

    Jiang S, Fang Y , Zhang H, Liu Q, Qi Y , Yang Y , Wang P (2026) Teacher agent: A knowledge distillation- free framework for rehearsal-based video incremen- tal learning. International Journal of Computer Vision 134(4):190

    2026

  26. [26]

    In: International Conference on Learning Rep- resentations, vol 2025, pp 34531–34582

    Kaul P, Ma C, Elezi I, Deng J (2025) From attention to activation: Unraveling the enigmas of large language models. In: International Conference on Learning Rep- resentations, vol 2025, pp 34531–34582

    2025

  27. [27]

    In: European Conference on Computer Vision (ECCV)

    Kembhavi A, Salvato M, Kolve E, Seo MJ, Hajishirzi H, Farhadi A (2016) A diagram is worth a dozen images. In: European Conference on Computer Vision (ECCV)

    2016

  28. [28]

    Advances in neural information processing sys- tems 29

    Kim JH, Lee SW, Kwak D, Heo MO, Kim J, Ha JW, Zhang BT (2016) Multimodal residual learning for vi- sual qa. Advances in neural information processing sys- tems 29

    2016

  29. [29]

    Pro- ceedings of Machine Learning and Systems 5:341–353

    Korthikanti V A, Casper J, Lym S, McAfee L, Ander- sch M, Shoeybi M, Catanzaro B (2023) Reducing acti- vation recomputation in large transformer models. Pro- ceedings of Machine Learning and Systems 5:341–353

    2023

  30. [30]

    Inter- national Journal of Computer Vision 133(1):31–64

    Liang J, He R, Tan T (2025) A comprehensive survey on test-time adaptation under distribution shifts. Inter- national Journal of Computer Vision 133(1):31–64

    2025

  31. [31]

    Neu- ral Computing and Applications 37(18):12631–12652

    Liang Y , Chen K (2025) Automatic test-time adapta- tion for heterogeneous contexts in meta-learning. Neu- ral Computing and Applications 37(18):12631–12652

    2025

  32. [32]

    In: European Conference on Computer Vision, Springer, pp 125–140

    Liu S, Zheng K, Chen W (2024) Paying more attention to image: A training-free method for alleviating halluci- nation in lvlms. In: European Conference on Computer Vision, Springer, pp 125–140

    2024

  33. [33]

    Liu Y , Duan H, Zhang Y , Li B, Zhang S, Zhao W, Yuan Y , Wang J, He C, Liu Z, et al (2024) Mmbench: Is your multi-modal model an all-around player? In: European conference on computer vision, Springer, pp 216–233

    2024

  34. [34]

    Knowledge-Based Systems p 115570

    Luo J, Ni Z, Liu W, Zhu X, Feng Y (2026) Dpc-tta: Dual perturbation and correction network guided test- time adaptation. Knowledge-Based Systems p 115570

    2026

  35. [35]

    In: Proceedings of Sinn und Be- deutung 23, vol 2, pp 107–124

    de Marneffe MC, Simons M, Tonhauser J (2019) The commitmentbank: Investigating projection in naturally occurring discourse. In: Proceedings of Sinn und Be- deutung 23, vol 2, pp 107–124

    2019

  36. [36]

    Scientific reports 7(1):10717

    Monsivais D, Bhattacharya K, Ghosh A, Dunbar RI, Kaski K (2017) Seasonal and geographical impact on human resting periods. Scientific reports 7(1):10717

    2017

  37. [37]

    arXiv preprint arXiv:180302999

    Nichol A, Achiam J, Schulman J (2018) On first- order meta-learning algorithms. arXiv preprint arXiv:180302999

    2018

  38. [38]

    International Journal of Computer Vision 134(6):311

    Ning KP, Ke HJ, Liu YY , Yao JY , Tian YH, Yuan L (2026) Sparse orthogonal parameters tuning for contin- ual learning. International Journal of Computer Vision 134(6):311

    2026

  39. [39]

    Advances in Neural Information Processing Systems 38:74741– 74778

    Noori M, Osowiechi D, Vargas Hakim G, Bahri A, Yaz- danpanah M, Dastani S, Beizaee F, Ayed I, Desrosiers C (2026) Test-time adaptation of vision-language models for open-vocabulary semantic segmentation. Advances in Neural Information Processing Systems 38:74741– 74778

    2026

  40. [40]

    In: Proceedings of the Com- puter Vision and Pattern Recognition Conference, pp 4430–4440

    Park U, Jeong S, Jang Y , Park GM, Ko JH (2025) Test-time fine-tuning of image compression models for multi-task adaptability. In: Proceedings of the Com- puter Vision and Pattern Recognition Conference, pp 4430–4440

    2025

  41. [41]

    In: International Conference on Learning Representations, vol 2024, pp 28412–28434

    Peng B, Tian Z, Liu S, Yang MC, Jia J (2024) Scalable language model with generalized continual learning. In: International Conference on Learning Representations, vol 2024, pp 28412–28434

    2024

  42. [42]

    In: International Conference on Learning Repre- sentations, vol 2024, pp 9695–9717

    Qin Y , Liang S, Ye Y , Zhu K, Yan L, Lu Y , Lin Y , Cong X, Tang X, Qian B, et al (2024) Toolllm: Facilitat- ing large language models to master 16000+ real-world apis. In: International Conference on Learning Repre- sentations, vol 2024, pp 9695–9717

    2024

  43. [43]

    Advances in Neural Information Process- ing Systems 37:55249–55285

    Qu Y , Zhang T, Garg N, Kumar A (2024) Recursive introspection: Teaching language model agents how to self-improve. Advances in Neural Information Process- ing Systems 37:55249–55285

    2024

  44. [44]

    URLhttps://qwen.ai/blog? id=qwen3.5

    Qwen Team (2026) Qwen3.5: Towards native mul- timodal agents. URLhttps://qwen.ai/blog? id=qwen3.5

    2026

  45. [45]

    In: The Eleventh Interna- tional Conference on Learning Representations

    Razdaibiedina A, Mao Y , Hou R, Khabsa M, Lewis M, Almahairi A (????) Progressive prompts: Continual learning for language models. In: The Eleventh Interna- tional Conference on Learning Representations

  46. [46]

    Nature Reviews Neuroscience 26(10):583–584

    Sadeh S, Clopath C (2025) The emergence of neuroai: bridging neuroscience and artificial intelligence. Nature Reviews Neuroscience 26(10):583–584

    2025

  47. [47]

    arXiv preprint arXiv:260119897

    Shenfeld I, Damani M, H ¨ubotter J, Agrawal P (2026) Self-distillation enables continual learning. arXiv preprint arXiv:260119897

    2026

  48. [48]

    Socher R, Perelygin A, Wu J, Chuang J, Manning CD, Ng A, Potts C (2013) Recursive deep models for seman- tic compositionality over a sentiment treebank. In: Pro- ceedings of the 2013 Conference on Empirical Meth- ods in Natural Language Processing, Association for Computational Linguistics, Seattle, Washington, USA, pp 1631–1642, URLhttps://www.aclweb.org...

    2013

  49. [49]

    arXiv preprint arXiv:230914763

    Song C, Han X, Zeng Z, Li K, Chen C, Liu Z, Sun M, Yang T (2023) Conpet: Continual parameter- efficient tuning for large language models. arXiv preprint arXiv:230914763

    2023

  50. [50]

    In: Proceedings of the 2019 Conference of the North American Chapter of 16 JX Liao et al

    Talmor A, Herzig J, Lourie N, Berant J (2019) Com- monsenseQA: A question answering challenge target- ing commonsense knowledge. In: Proceedings of the 2019 Conference of the North American Chapter of 16 JX Liao et al. the Association for Computational Linguistics: Human Language Technologies, V olume 1 (Long and Short Pa- pers), Association for Computati...

    2019

  51. [51]

    Advances in Neural Information Processing Systems 37:22624– 22656

    Thrampoulidis C (2024) Implicit optimization bias of next-token prediction in linear models. Advances in Neural Information Processing Systems 37:22624– 22656

    2024

  52. [52]

    Sci- ence advances 6(16):eaay2631

    Udrescu SM, Tegmark M (2020) Ai feynman: A physics-inspired method for symbolic regression. Sci- ence advances 6(16):eaay2631

    2020

  53. [53]

    Advances in neural information process- ing systems 30

    Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, Kaiser Ł, Polosukhin I (2017) Attention is all you need. Advances in neural information process- ing systems 30

    2017

  54. [54]

    arXiv preprint arXiv:240912191

    Wang P, Bai S, Tan S, Wang S, Fan Z, Bai J, Chen K, Liu X, Wang J, Ge W, Fan Y , Dang K, Du M, Ren X, Men R, Liu D, Zhou C, Zhou J, Lin J (2024) Qwen2-vl: Enhanc- ing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:240912191

    2024

  55. [55]

    In: Proceed- ings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers), pp 13484–13508

    Wang Y , Kordi Y , Mishra S, Liu A, Smith NA, Khashabi D, Hajishirzi H (2023) Self-instruct: Aligning language models with self-generated instructions. In: Proceed- ings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers), pp 13484–13508

    2023

  56. [56]

    In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp 11548–11565

    Wu T, Yuan W, Golovneva O, Xu J, Tian Y , Jiao J, Weston JE, Sukhbaatar S (2025) Meta-rewarding lan- guage models: Self-improving alignment with llm-as-a- meta-judge. In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp 11548–11565

    2025

  57. [57]

    co/datasets/xai-org/RealworldQA, ac- cessed: 2025-04-26

    xAI (2024) Realworldqa: A benchmark for real-world spatial understanding.https://huggingface. co/datasets/xai-org/RealworldQA, ac- cessed: 2025-04-26

    2024

  58. [58]

    In: International Conference on Learning Representa- tions, vol 2024, pp 21875–21895

    Xiao G, Tian Y , Chen B, Han S, Lewis M (2024) Ef- ficient streaming language models with attention sinks. In: International Conference on Learning Representa- tions, vol 2024, pp 21875–21895

    2024

  59. [59]

    URLhttps://arxiv.org/abs/2505

    Xiaomi LCT (2025) Mimo: Unlocking the reasoning potential of language model – from pretraining to post- training. URLhttps://arxiv.org/abs/2505. 07608,2505.07608

    Pith/arXiv arXiv 2025

  60. [60]

    IEEE Transactions on Knowledge & Data Engineering 38(07):4674–4693, DOI 10.1109/TKDE.2026.3675811

    Xu K, Wang L, Zhang H, Yin B (2026) Self- Knowledge Distillation and Its Application: A Survey . IEEE Transactions on Knowledge & Data Engineering 38(07):4674–4693, DOI 10.1109/TKDE.2026.3675811

    arXiv 2026

  61. [61]

    IEEE Transactions on Pattern Analysis and Machine In- telligence

    Xu L, Xie H, Qin SJ, Tao X, Wang FL (2026) Parameter-efficient fine-tuning methods for pretrained language models: A critical review and assessment. IEEE Transactions on Pattern Analysis and Machine In- telligence

    2026

  62. [62]

    Advances in Neural In- formation Processing Systems 38:23706–23741

    You C, Wu K, Jia Z, Chen L, Bhojanapalli S, Guo J, Evci U, Wassenberg J, Netrapalli P, Willcock J, et al (2026) Spark transformer: Reactivating sparsity in transformer ffn and attention. Advances in Neural In- formation Processing Systems 38:23706–23741

    2026

  63. [63]

    arXiv preprint arXiv:240615765

    Yu Z, Wang Z, Fu Y , Shi H, Shaikh K, Lin YC (2024) Unveiling and harnessing hidden attention sinks: En- hancing large language models without training through attention calibration. arXiv preprint arXiv:240615765

    2024

  64. [64]

    In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pp 9556–9567

    Yue X, Ni Y , Zhang K, Zheng T, Liu R, Zhang G, Stevens S, Jiang D, Ren W, Sun Y , et al (2024) Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pp 9556–9567

    2024

  65. [65]

    arXiv preprint arXiv:221008340

    Zador A, Escola S, Richards B, Bengio Y , Boahen K, Botvinick M, Chklovskii D, Churchland A, Clopath C, DiCarlo J, et al (2022) Toward next-generation artificial intelligence: Catalyzing the neuroai revolution. arXiv preprint arXiv:221008340

    2022

  66. [66]

    In: International Conference on Learning Representations, vol 2024, pp 42411–42430

    Zhang Q, Singh C, Liu L, Liu X, Yu B, Gao J, Zhao T (2024) Tell your model where to attend: Post-hoc atten- tion steering for llms. In: International Conference on Learning Representations, vol 2024, pp 42411–42430

    2024

  67. [67]

    arXiv preprint arXiv:240910790

    Zhang Q, Yu X, Singh C, Liu X, Liu L, Gao J, Zhao T, Roth D, Cheng H (2024) Model tells itself where to attend: Faithfulness meets automatic attention steering. arXiv preprint arXiv:240910790

    2024

  68. [68]

    arXiv preprint arXiv:260401193

    Zhang R, Bai RH, Zheng H, Jaitly N, Collobert R, Zhang Y (2026) Embarrassingly simple self- distillation improves code generation. arXiv preprint arXiv:260401193

    2026

  69. [69]

    Journal of Machine Learning Research 25(194):1–74

    Zhang SQ, Chen JY , Wu JH, Zhang G, Xiong H, Gu B, Zhou ZH (2024) On the intrinsic structures of spiking neural networks. Journal of Machine Learning Research 25(194):1–74

    2024

  70. [70]

    arXiv preprint arXiv:260118734

    Zhao S, Xie Z, Liu M, Huang J, Pang G, Chen F, Grover A (2026) Self-distilled reasoner: On-policy self- distillation for large language models. arXiv preprint arXiv:260118734

    2026

  71. [71]

    In: Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pp 8363–8371

    Zhou DW, Sun HL, Ning J, Ye HJ, Zhan DC (2024) Continual learning with pre-trained models: a survey. In: Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pp 8363–8371

    2024

  72. [72]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference, pp 1624–1633

    Zhu L, Ji D, Chen T, Xu P, Ye J, Liu J (2025) Ibd: Al- leviating hallucinations in large vision-language mod- els via image-biased decoding. In: Proceedings of the Computer Vision and Pattern Recognition Conference, pp 1624–1633

    2025