---
title: Adaptive Originality Filtering (AOF)
url: https://www.emergentmind.com/topics/adaptive-originality-filtering-aof
type: topic
---

# Adaptive Originality Filtering (AOF)

Searching arXiv for the specified paper to ground the article in the cited source.
arXiv search query: 2508.18709
Adaptive Originality Filtering (AOF) is a lightweight, prompting-only framework for steering large language models (LLMs) toward more creative, culturally grounded, and lexically novel riddles in multiple languages. It interposes a rejection-sampling loop around a specially structured riddle prompt and enforces three orthogonal constraints: Semantic Novelty, Lexical Novelty, and Cross-Lingual Fidelity. In multilingual riddle generation, its stated objectives are to reduce rote reuse of commonplace or memorized riddles, increase lexical diversity, and maintain cultural fluency and semantic alignment across two languages. The framework is introduced in "Filtering for Creativity: Adaptive Prompting for Multilingual Riddle Generation in LLMs" [2508.18709].

## 1. Conceptual definition and problem setting

Multilingual riddle generation challenges large language models to balance cultural fluency with creative abstraction. Standard prompting strategies—zero-shot, few-shot, and chain-of-thought—tend to reuse memorized riddles or perform shallow paraphrasing. AOF is proposed in response to that failure mode as a prompting framework that filters redundant generations using cosine-based similarity rejection, while enforcing lexical novelty and cross-lingual fidelity [2508.18709].

The framework is explicitly defined as a rejection-sampling wrapper around standard prompts. Its three constraints are orthogonal in the sense stated by the source: semantic novelty targets embedding-level similarity to a reference corpus of known riddles; lexical novelty targets formulaic or too-common riddles by banning generic answers and enforcing diverse grammatical templates; and cross-lingual fidelity targets semantic alignment and answer preservation in bilingual generation. In that formulation, AOF is not a task-specific fine-tuning method. It is a prompting-only mechanism for steering generation, although the reported study also considers a fine-tuned GPT-4o condition.

AOF is situated specifically within riddle generation rather than as a general creativity framework. The source states that generalizing to other creative tasks is untested. A plausible implication is that AOF should be understood first as a domain-specific methodology for multilingual riddle production, with broader applicability remaining an empirical question rather than an established property.

## 2. Rejection-sampling architecture

AOF takes as input a model $M$, an AOF prompt template $P_{\text{AOF}}$, a reference set of known riddles $D = \{r_i\}$, an embedding function $\phi(\cdot)$, a cosine-similarity threshold $\theta$, a banned-answer list $A$, a cross-lingual fidelity check function $\mathrm{CL\_check}(r_{\text{gen}})$, and a maximum number of attempts $k$ [2508.18709]. The generation process is organized as an iterative filter:

```text
function GENERATE_RIDDLE_AOF(M, P_AOF, D, φ, θ, A, CL_check, k):
  for attempt in 1..k:
    r_gen ← M.generate(P_AOF)
    # 1. Semantic novelty check
    S ← max_{r_i ∈ D} cosine(φ(r_gen), φ(r_i))
    if S ≥ θ:
      continue  # too similar, reject
    # 2. Lexical novelty check
    if extract_answer(r_gen) ∈ A:
      continue  # banned generic answer
    # 3. Cross-lingual fidelity (bilingual setting)
    if not CL_check(r_gen):
      continue  # translation or semantic drift
    return r_gen
  return None  # give up after k attempts
```

This architecture makes the acceptance criterion conjunctive: a candidate riddle must pass all three filters before it is returned. The reference set $D$ is drawn from BiRdQA training splits of known riddles. The banned-answer list is given as $\{\text{“shadow”, “time”, “echo”, “fire”, “breath”, “wind”, “silence”}\}$, and the cosine-similarity threshold is set to $\theta = 0.75$ in all experiments. The maximum number of retries is typically $3$–$5$ per riddle slot.

Operationally, the framework delegates candidate production to the base LLM and places originality control outside the model in a lightweight wrapper. That design choice is central to the paper’s claim that semantic rejection can guide culturally grounded, creative generation without task-specific fine-tuning.

## 3. Formal criteria for novelty and bilingual consistency

The semantic-rejection mechanism is defined through cosine similarity over embeddings. For two embedding vectors $u, v \in \mathbb{R}^d$,

$$
\mathrm{sim}(u,v) = \frac{u \cdot v}{\|u\| \cdot \|v\|}
$$

For a candidate riddle $r_{\text{gen}}$ and reference set $D$,

$$
S(r_{\text{gen}}, D) = \max_{r_i \in D} \mathrm{sim}(\phi(r_{\text{gen}}), \phi(r_i))
$$

A candidate is accepted only if $S(r_{\text{gen}}, D) < \theta$, with $\theta = 0.75$ in all experiments [2508.18709].

Lexical novelty is quantified with Distinct-$n$, defined for a corpus $C$ of generated riddles as

$$
\mathrm{Distinct}\text{-}n = \frac{\text{Total number of distinct } n\text{-grams in } C}{\text{Total number of } n\text{-grams in } C}
$$

Higher Distinct-$n$ indicates more lexical diversity. Redundancy is quantified with Self-BLEU. Given $C = \{c_1, \ldots, c_N\}$,

$$
\mathrm{Self\text{-}BLEU} = \frac{1}{N} \sum_{i=1}^{N} \mathrm{BLEU}(c_i, C \setminus \{c_i\})
$$

Lower Self-BLEU indicates less redundancy.

Cross-lingual fidelity is checked after the novelty filters. The English and foreign-language versions must remain semantically aligned and share the same answer. In practice, this was operationalized via a Cross-Lingual BERTScore or embedding-based similarity threshold, with the example threshold stated as BERTScore F1 $> 0.80$ between the two parts. The AOF prompt also explicitly instructs the model to “preserve ambiguity” and “use the same answer in both languages.” This coupling of semantic novelty with bilingual answer consistency is one of the framework’s defining features.

## 4. Experimental configuration

The experimental setup covers three models—GPT-4o (OpenAI), LLaMA 3.1 (Meta), and DeepSeek R1 (R1)—and four language pairs: English⇄Arabic, English⇄Chinese, English⇄Japanese, and English⇄French [2508.18709]. The prompt designs compared are Zero-Shot, Few-Shot, Chain-of-Thought, Adversarial, and AOF. Few-Shot uses $3$–$5$ examples, while Chain-of-Thought is described as CoT with stepwise reasoning and Adversarial as distractor-rich prompts.

Generation hyperparameters are fixed across all generation calls: temperature $= 0.7$ and max tokens $= 3000$. The embedding function is exemplified by all-MiniLM-L6-v2. Automatic metrics are Self-BLEU ($n=2$), Distinct-2, Cross-Lingual BERTScore, token length, and syntactic validity. The source also reports a sensitivity ablation over $\theta$, concluding that $0.75$ best balances Self-BLEU vs. Distinct-2.

This setup isolates AOF from changes in decoding policy and model architecture. The paper correspondingly notes that interactions between AOF and decoding hyperparameters remain to be explored, since prompting parameters were held fixed.

## 5. Reported empirical behavior

The reported results separate an excerpt for pretrained models under AOF from a fine-tuned GPT-4o condition. For pretrained models under AOF, the source reports the following excerpt from the table “performance_bleu_distinct”: English–Chinese (GPT-4o) with Self-BLEU $= 0.280$ and Distinct-2 $= 0.869$, and English–Japanese (GPT-4o) with Self-BLEU $= 0.483$ and Distinct-2 $= 0.697$ [2508.18709].

For fine-tuned GPT-4o with AOF, the source reports:

| Language pair | Self-BLEU | Distinct-2 |
|---|---:|---:|
| English–Arabic | 0.260 | 0.893 |
| English–Chinese | 0.163 | 0.934 |
| English–Japanese | 0.177 | 0.915 |
| English–French | 0.273 | 0.856 |

The paper states that AOF yields the lowest Self-BLEU, up to $50 \%$ lower, and the highest Distinct-2 across most language pairs, signaling both reduced redundancy and increased lexical variety. The Japanese setting is emphasized: fine-tuned AOF achieved Self-BLEU $0.177$ versus $0.431$ for Zero-Shot, and Distinct-2 $0.915$ versus $0.752$. The abstract likewise highlights that AOF-enhanced GPT-4o achieves $0.177$ Self-BLEU and $0.915$ Distinct-2 in Japanese, interpreting these values as evidence of improved lexical diversity and reduced redundancy compared to other prompting methods and language pairs.

The study’s summary claim is therefore not merely that AOF increases novelty metrics, but that it improves the balance between creativity, operationalized through Distinct-$n$, and non-redundancy, operationalized through Self-BLEU.

## 6. Limitations, boundary conditions, and interpretation

The reported limitations are explicit. Dataset scope is limited to BiRdQA, described as primarily English and Chinese riddles; generalizing to other creative tasks is untested. The MiniLM-based semantic filter emphasizes surface embedding distances and may miss deeper paraphrases in low-resource languages. Prompting parameters were held fixed, so interactions between AOF and decoding hyperparameters remain unexplored. Cross-lingual fidelity was enforced by embedding-based checks, but finer-grained translation quality or cultural idiom alignment was not fully ablated. The threshold ablation showed an optimal novelty threshold at $\theta = 0.75$: lower values over-reject, hurting diversity, while higher values under-reject, increasing redundancy [2508.18709].

These limitations delimit the appropriate interpretation of AOF. The framework is supported as a modular rejection-sampling wrapper around standard prompts that systematically enforces novelty and bilingual consistency within multilingual riddle generation. The source further states that it requires no model fine-tuning when used in pretrained LLMs, yet, when combined with light multiple-choice fine-tuning, delivers state-of-the-art balance of creativity and non-redundancy in multilingual riddle generation. A plausible implication is that the principal contribution lies less in a new generative objective than in a practically deployable control layer for filtering memorization, formulaicity, and bilingual drift under fixed prompting conditions.

Source: https://www.emergentmind.com/topics/adaptive-originality-filtering-aof