acceptodds
Under review as a conference paper at ICLR 2027

CANVAS: Adjustable Creativity Control in Large Language Models via Sparse Neuron Scaling

Abstract

Large language models often produce homogeneous responses to open-ended creative requests, yet increasing variation alone does not ensure meaningful novelty. We study adjustable creativity control without retraining, addressing two challenges: distinguishing novelty from uncontrolled deviation and identifying neurons whose amplification improves creative generation. We introduce CANDLE, an evaluation framework that measures semantic and lexical deviation from the unmodified model's fixed, prompt-specific greedy response under explicit quality constraints. Building on this framework, CANVAS ranks feed-forward neurons by the estimated first-order effect of scaling on a creative-token objective, filters potentially destabilizing targets, and scales a sparse subset of weights. Intervention strength is continuously adjustable, and fixed-strength edits require no additional inference-time computation. Experiments across three model families and three creative-task categories show improved empirical novelty–quality trade-offs. CANDLE's novelty metric achieves agreement scores of with LLM judges and with human annotators. At a scaling strength of , CANVAS increases novelty scores by on Qwen and on Gemma, with no observed decline in mean quality. Ablations support the contributions of intervention-aligned attribution and stability safeguards. These results establish sparse neuron scaling as a practical approach to adjustable, quality-constrained creative generation.

Primary area:
interpretability and explainable AI
Keywords:
Large language models, creativity evaluation, controllable generation, neuron-level intervention, sparse weight scaling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_mc24OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces CANDLE and CANVAS, with CANDLE as an evaluation framework for a model's creativity under a relative novelty score against the model's own default generations. It works by comparing a weighted average of semantic deviation and lexical deviation against the default generation (the model without any modifications). CANVAS is a steering framework applied to the model's weights to steer the behavior of outputting more creative generations. It works by generating responses from the model and judging those according to some rubrics on validity and novelty, and then the authors used the Creativity Attribution Score (an activation-gradient product value that indicates which neurons can be adjusted to make the model more creative) to apply a steering scaler to the FFN down projection matrix of top k neurons for steering the model's creativity. The paper also introduces two safeguard mechanisms to prevent model from collapsing. The paper tested three families of models on their methods and compared against a broad range of ways of steering the model's creativity, and shows their method is usually the best under their evaluation framework. Lastly, they show a large human-AI alignment on the creativity judgement of the final results.

Critical strengths

The strength is that the paper proposes a steering method that is easy to use, and it provides quick model updates without substantial training or inference time, while preserving the model's general capabilities quite well. This is important since it is efficient and easy to apply and has some possible practical uses if one wants to have a more creative model.

Critical weaknesses

The evaluation is a bit constrained, and a more holistic evaluation of creativity (under different criteria or task types) might be better to demonstrate the effectiveness of their method. For instance, try creativity in different domains such as math. From my understanding. CANDLE measures more of the relative gains of the model to itself, which may omit dimensions of creativity in general. In addition, CANVAS depends on GPT-4o for quality filtering and creative-span annotation, and thus the selected neurons may reflect the annotators' preferences. If the authors can show that with other creativity-related metrics, CANVAS can still perform well, I will consider to raise my score.

Questions for authors
  1. Can authors show that with other creativity-related metrics, CANVAS can still perform well across different kinds of creativity-related tasks that are relatively more out of domain compared to the tasks being trained on? For instance, can this method make the model produce more creative math solutions or poems, while maintaining correctness or format constraints?
  2. Can the author show if the alpha score lower than 1 means less creative outputs? I wonder if there is an interpretable understanding of the alpha score and if it is well calibrated as well. This is important since if the method is applied to general cases, we want to ensure the alpha is well-calibrated under different creativity requirements.
Reviewer_GR8zOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a framework for evaluating creativity in LLMs (named CANDLE), as well as a weight-space steering method named CANVAS that uses the CANDLE framework to improve the creativity metrics introduced in CANDLE. The CANDLE evaluation framework involves measuring semantic and lexical deviation from a base model's fixed response, conditioned on that deviation being useful and appropriate. Empirical results reveal that this method improves on open-weight models relative to other steering baselines, and a small human study confirms that humans prefer the creativity of CANVAS models over baseline models.

Critical strengths

Addresses an important problem and is well-motivated by prior methodological works. This paper addresses an important problem with existing knowledge and skill localization methods like H-Neuro, CreativityNeuro, etc. that find weights associated with a particular behavior; however, association is not the same as causality, and steering associated weights does not guarantee that this will improve the behavior.

Critical weaknesses

Soundness concerns about baselines. I have concerns about the settings used among baselines compared against in this work, and the omission of key parameters in baseline methods.

In Section C.2, the paper states “every baseline exposes a strength parameter, swept so that each method traces a full curve on the creativity-quality Pareto plane.”

However, several of the baselines have multiple parameter choices beyond a strength parameter that are not mentioned at all in the paper. For example, contrastive activation addition (CAA) requires

  • (1) inputs that correspond to a target behavior (+) and its opposite (-), in order to find a direction in the residual stream corresponding to that behavior by performing subtraction between the (+) and (-) vectors
  • (2) as well as a specification of which layers to intervene, though this is never mentioned in the paper.

Similarly, it appears CreativityNeuro also requires

  • (1) a dataset of contrastive prompts, like CAA, to localize creativity-specific weights
  • (2) as well as the ρ\rho parameter that details the selectivity of its mask, though Table 7 of the appendix completely omits this hyperparameter, and the paper never mentions this ρ\rho parameter at all.

Lastly, the authors search a much wider range of strength parameters in Table 7 for their own method (5-6 parameter choices) than they do for the other steering baselines (usually only 3). This is not a fair comparison, and I have concerns that the baselines have been under-tuned.

Mode deviation novelty (MDN) and empirical evaluation of CANVAS are not well-grounded in prior creativity evaluation frameworks. The same metrics (MDN) and datasets (InfinityChat) that are used to identify neurons in the method are also used to evaluate its efficacy. It is great and much appreciated that the authors conducted a human evaluation, but this evaluation is limited and only done on Qwen3-8B relative to the baseline model, and no human ever compares CANVAS with another steering method, when CAA is shown to outperform CANVAS in Figure 15(a).

Moreover, the OOD prompts evaluated in Appendix Section D.5. are also drawn from InfinityChat, only MDN and GPT-4o pairwise preferences are reported as metrics, and there are no established creativity tests like the DAT, AUT, etc. that verify this method improves “creativity” in general rather than MDN on a limited set of prompts. It is crucial to see whether the method generalizes beyond questions in InfinityChat, and I strongly urge the authors to evaluate whether their method works on other well-known creativity tasks.

More broadly, a central focus of this paper is about improving creativity in LLMs, but the work does not engage with the rich literature in cognitive science and psychology (Runco, Boden, Simonton, etc.) about how to evaluate creative products at more than a superficial level. Moreover, it is unclear how this measure of novelty differs from other ways of calculating novelty. Novelty is typically defined as the distance of an artifact relative to the existing class of artifacts (Maher; 2010) [https://dl.acm.org/doi/10.5555/1854969.1854977], but here it is defined as deviation from a baseline response, which many experts on creativity might not consider an appropriate way to operationalize novelty. For rebuttals, the authors could consider another definition of novelty introduced in [https://arxiv.org/abs/2509.02534] that assesses novelty/diversity relative to a set of existing responses, rather than just the baseline response.

Questions for authors
  • The phrase “uncontrolled deviation” is used several times throughout the paper. What exactly does this mean? Deviation from what?
  • Table 7 is described as “recovered from the saved experimental results.” What does this mean, and why is it mentioned in the paper?
  • In Section 5.1, the paper states that “random scaling also increases creativity in some settings.” Where were these findings made? If they were in this work, the section (possibly in appendix) that found this should be referenced; if they were in other work, the other work should be referenced.
  • In Equation 6, sa,is_{a, i} and sb,is_{b, i} are never defined. Can this be fixed?
  • In Equation 2, what symbol is being used on the gate and pass rate? Is this supposed to be the indicator function? (i.e, I\mathbb{I})
  • One reference is duplicated (Turner 2024a and 2024b are the same). Can this be fixed?
Reviewer_qkGyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CANDLE, an evaluation framework that measures creativity through semantic and lexical deviation from the model's greedy response under quality constraints. Building on this metric, the authors introduce CANVAS, which identifies neurons associated with creative generation using gradient-based attribution and scales their weights to control creativity without retraining. Experiments across three model families show improved novelty-quality trade-offs compared to several baselines, supported by ablations and human and LLM evaluations.

Critical strengths
  • I think it is a practical approach to creativity control. The proposed method is simple and efficient, allowing adjustable control through sparse weight scaling without additional inference-time computation.
  • The authors evaluate multiple model families and compare against several existing approaches. The ablations and additional evaluations provide useful evidence for the effectiveness of the method.
Critical weaknesses
  • The evaluation metric may not adequately capture meaningful creativity. CANDLE primarily measures deviation from a single greedy response, which does not necessarily correspond to creativity or originality. The quality gate ensures basic appropriateness but does not distinguish genuinely creative outputs from merely different ones. This concern is particularly important because the same metric also guides the construction of the data used for neuron selection. Although the authors conduct human evaluations, these primarily compare CANVAS against the unmodified model, rather than stronger baselines. Thus, it remains unclear whether CANVAS improves meaningful creativity beyond what existing approaches achieve.
  • The method has limited novelty. The core approach is closely related to CreativityNeuro, which also identifies and scales creativity-associated neurons. The main methodological differences appear to be the use of established gradient-based attribution and additional stability filters. While these modifications improve empirical performance, I find the conceptual contribution relatively incremental.
Questions for authors
  • Can the authors provide direct human evaluations comparing CANVAS against strong baselines such as CreativityNeuro and CAA? This would help establish whether the improvement in CANDLE corresponds to meaningful creativity gains rather than greater deviation from the default response.
  • How robust are the conclusions to alternative definitions of creativity that do not rely on a single greedy response? In particular, could the authors evaluate originality relative to a broader distribution of model outputs?
  • Can the authors further clarify the conceptual contribution over CreativityNeuro and isolate how much of the improvement comes from the attribution method versus the stability safeguards?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.