TLDR: A new decoding strategy called Avoidance Decoding helps Large Language Models generate more diverse and less repetitive multi-branch stories. It works by penalizing tokens that are too similar to previously generated outputs, using both concept-level and narrative-level similarity measures. This method significantly boosts output diversity, reduces repetition, and activates more of the model’s intrinsic creative capacity without additional training.
Large Language Models (LLMs) have shown incredible capabilities in generating text, but they often struggle with creativity, especially when tasked with generating multiple variations from the same input. This can lead to repetitive and monotonous outputs, a significant challenge in creative tasks like story generation.
Researchers Kyeongman Park, Nakyeong Yang, and Kyomin Jung from Seoul National University have introduced a novel decoding strategy called Avoidance Decoding to tackle this very problem. Their method aims to encourage more diverse multi-branch stories by preventing LLMs from generating content too similar to what they’ve already produced.
How Avoidance Decoding Works
The core idea behind Avoidance Decoding is to modify the probabilities of tokens an LLM might choose next. It does this by applying a penalty to tokens that are similar to previously generated outputs. This penalty is not static; it adaptively balances two key similarity measures:
- Concept-level Similarity Penalty (CSP): In the early stages of story generation, this penalty is prioritized. Its goal is to diversify the initial ideas and concepts of the story branches, ensuring they start off on distinct paths.
- Narrative-level Similarity Penalty (NSP): As the story progresses and becomes longer, this penalty gains more emphasis. It focuses on ensuring that the plot development remains natural yet diverse, preventing the narratives from converging too much.
By combining these two penalties in a hybrid approach, Avoidance Decoding effectively steers the LLM to explore a wider range of creative possibilities. It doesn’t require any additional training for the LLM or complex stochastic sampling methods.
Impressive Results
The researchers conducted extensive experiments using various LLMs, including Mistral 7B, Llama 3B, Llama 8B, and Qwen 7B, across different story prompt datasets. The results are compelling:
- The method achieved up to 2.6 times higher output diversity compared to strong baseline methods.
- It reduced repetition in generated texts by an average of 30%.
- Crucially, it effectively mitigated text degeneration, a common issue where models start producing incoherent or nonsensical text when pushed for diversity.
Beyond quantitative metrics, the study also revealed that Avoidance Decoding activates a broader range of neurons within the LLM. This suggests that the method isn’t just introducing superficial variations but is actually tapping into the model’s intrinsic creative capacity.
For instance, when generating two stories from the same prompt about a blind date, the method could produce one story with a cheerful, bustling city setting and another with a somber, rainy atmosphere, complete with different emotional tones and plot points. This qualitative example clearly demonstrates the method’s ability to foster both conceptual and narrative divergence.
Also Read:
- Top-H Decoding: A Smarter Way for LLMs to Balance Creativity and Coherence
- Ensuring Language Models Reason for the Right Reasons
Looking Ahead
While Avoidance Decoding shows significant promise, the researchers acknowledge a limitation: increased decoding time, especially as the number of previously generated negative samples grows. Future work could explore solutions like storing only a fixed-size window of recent outputs to manage computational overhead.
This research marks a significant step forward in enhancing the creative capabilities of LLMs, particularly for tasks requiring diverse and engaging multi-branch narratives. You can read the full research paper here.
