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

🕒 Published (v1): 2026-07-15 02:43 UTC · Source: Arxiv · link

Why this paper was selected

Phase-transition self-improvement phenomenon; Tianwei Zhang (NTU); actionable finetuning recipe

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Enlightenment is a training-free, data-free, gradient-free post-tuning paradigm that modifies shortcut connections (residual pathways) in pre-trained models rather than their weights. For VLMs, it applies a single shared scalar \(\eta \in (0,1]\) to every residual connection in decoder layers; for LLMs, it injects cross-head shortcuts aggregating head-0 outputs into all other heads. The method operates in a novel "retrospection" phase between training and inference, inspired by the neuroscience of sudden insight.

Problem

Existing training-free finetuning methods exclusively manipulate attention weights (e.g., redistributing BOS/"attention sink" weight), leaving the architectural/structural dimension of models unexplored. No prior method operates entirely without data, gradients, and weight updates while also modifying model structure rather than parameters.

Method

The unified principle is: scale existing shortcuts or inject new ones, treating the scaling factor as a structural hyperparameter rather than a trainable weight.

For VLMs (decoder-only layers, both attention and FFN sub-blocks): replace the standard residual update with a scalar-damped variant: $\(x_{\ell+1} = x_\ell + \eta \cdot F_\ell(\text{LN}(x_\ell)), \quad \eta \in (0, 1]\)$ A single shared \(\eta\) suppresses overly active sublayers, reducing visual noise while preserving the primary residual stream. \(\eta = 1\) is the unmodified baseline; \(\eta = 0.5\) halves sublayer contribution uniformly.

For LLMs (multi-head attention): (i) Scale each head's BOS attention weight by a per-head factor \(\gamma_h\) fitted via a 3rd-order Chebyshev polynomial \(\gamma_h = \beta \cdot T_d(\hat{a}_0^{(h)}; \mathbf{c})\), where \(\hat{a}_0^{(h)}\) is the normalized mean BOS attention for head \(h\); (ii) insert new cross-head shortcuts injecting a fraction \(\alpha\) of head-0's output into every other head: $\(\tilde{\mathbf{O}} = [\tilde{\mathbf{A}}_0\mathbf{V}_0,\ \alpha\tilde{\mathbf{A}}_0\mathbf{V}_0 + \tilde{\mathbf{A}}_1\mathbf{V}_1,\ \ldots,\ \alpha\tilde{\mathbf{A}}_0\mathbf{V}_0 + \tilde{\mathbf{A}}_H\mathbf{V}_H]\)$ Default hyperparameters: polynomial degree \(d=3\), \(\beta=1\), \(\alpha=0.1\).

Key Contributions

• Retrospection paradigm: a novel operating phase (post-training, pre-inference) that modifies shortcuts without data, gradients, or weight updates — distinct from continual learning, self-distillation, test-time adaptation, and prior training-free finetuning.
• VLM instantiation: scalar-modulated residual damping applied uniformly across all decoder layers, requiring only one hyperparameter \(\eta\) and one multiply-per-layer overhead.
• LLM instantiation: Chebyshev-polynomial-driven adaptive BOS scaling combined with head-mixing shortcuts that expand representational capacity without adding weights.
• Empirical breadth: evaluated on three VLM variants and three LLM variants (Qwen3.5-9B-Base, DeepSeek-R1-Distill-Qwen-14B, Mimo-7B-RL) across multimodal and reasoning benchmarks.

Results

The experimental section is truncated in the provided text; no specific benchmark numbers or baseline comparisons are available in the supplied excerpt. The paper claims "remarkable performance improvements across diverse benchmarks and models" but quantitative results are not extractable from this text.

Limitations

• The task-dependent scaling factor \(\beta\) requires per-task tuning, partially breaking the "data-free" claim in practice (even a small validation signal is needed to set \(\beta\)).
• The neuroscience analogy is explicitly stated as a "loose conceptual analogy" with no claimed direct biological basis — the motivation is heuristic, not mechanistic.
• Hyperparameter sensitivity of \(\eta\) (VLM) and \(\alpha\) (LLM) requires grid search; claimed insensitivity is asserted but not fully demonstrated in the provided text.
• The cross-head shortcut design for LLMs always routes through head 0, which may be suboptimal if head 0 is not representationally rich for a given task or model architecture.
• Experiments cover only Qwen2-family and Qwen3-family backbones; generalization to architectures with different attention designs (e.g., GQA, sliding-window attention) is unverified.

Relevance to Vision-Language Models

Enlightenment's VLM-specific module directly targets a known failure mode of multimodal decoder models: overly active sublayers amplifying visual noise during inference. The scalar-modulated residual (\(\eta < 1\)) acts as a post-hoc regularizer on the LLM decoder component of VLMs, offering a complementary tool to attention-reweighting methods that have been used to reduce hallucinations (e.g., redistributing attention toward visual tokens). Unlike adapter-based or LoRA-style VLM tuning, this method introduces zero new parameters and requires no image-text training pairs, making it applicable to any frozen VLM with a standard transformer decoder. For researchers tracking VLMs, this work opens a structural (rather than parametric) axis of post-deployment adaptation that has been largely unexplored.