PaperlayerView exact revision on arXiv

Source-linked AI summary

Optimizing Diversity and Quality through Base-Aligned Model Collaboration

Yichen Wang, Chenghao Yang, Tenghao Huang, Muhao Chen, Jonathan May, Mina Lee

arXiv:2511.05650v2cs.CLcs.AIcs.LG

TL;DR

Alignment improves output quality but reduces diversity across generations, especially in open-ended tasks. BACO routes token generation between base and aligned models; across three tasks, its best router improves diversity and quality over strong inference-time baselines.

  • Problem

    Alignment can make outputs highly similar across generations, leaving a diversity-quality trade-off in open-ended generation.

    Source: §1 · p. 1

  • Method

    BACO dynamically routes token generation between a base model and its aligned counterpart using uncertainty- and content-based signals.

    Source: §1 · pp. 2–3, §3.1 · p. 4

  • Results

    Across three open-ended generation tasks, BACO’s best router achieves a 21.3% improvement over the strongest inference-time baselines, consistently across tasks and metrics and supported by human evaluations.

    Source: §1 · p. 2

  • Takeaways & Limitations

    An adjustable routing threshold and complementary strategies provide a controllable way to balance diversity and quality during generation.

    Source: §1 · p. 2

    Collaboration incurs additional FLOPs per token, and the prototype is not yet perfectly engineered for speed.

    Source: Limitations · p. 16

Abstract

from arXiv · show

Alignment has greatly improved large language models (LLMs)' output quality at the cost of diversity, yielding highly similar outputs across generations, especially in open-ended generation tasks. We propose Base-Aligned Model Collaboration (BACo), an inference-time token-level model collaboration framework that dynamically combines a base LLM with its aligned counterpart to optimize diversity and quality. Using uncertainty and content-based signals, BACo employs routing strategies to determine, at each token, which model to decode from. Prior diversity-promoting methods often improve diversity at the expense of quality or require expensive decoding or post-training. In contrast, BACo achieves both high diversity and quality post hoc within a single pass, while offering strong controllability. We introduce a family of effective routing strategies and evaluate them across three open-ended generation tasks with 13 diversity and quality metrics. BACo consistently surpasses state-of-the-art inference-time baselines. With our best router, BACo achieves a 21.3% joint improvement in diversity and quality, which is further supported by human evaluations. Overall, our results demonstrate that collaboration between base and aligned models provides an effective and controllable mechanism for optimizing the diversity-quality trade-off.

1. Introduction

Alignment improves LLM output quality but reduces diversity across generations, a tension that limits open-ended generation. BACO combines base and aligned models through token-level routing, and its best router outperforms inference-time baselines across tasks.

1. Introduction

  • Alignment can make repeated outputs highly similar, undermining open-ended generation through formulaic language, diminished creativity, and suppressed ideation.The paper motivates methods that improve diversity in aligned LLMs.Source: §1 · p. 1
  • Prior diversity-promoting approaches either modify model output distributions, potentially compromising alignment properties, or use inference-time techniques that may require multiple passes and degrade quality.The paper identifies these limitations as motivation for a method that balances diversity and quality at inference time.Source: §1 · p. 1
  • BACO dynamically routes token generation between a base model, which supplies diversity, and its aligned counterpart, which supplies quality, in a single pass without fine-tuning or prompting.The framework uses lightweight token-level routing strategies.Source: §1 · pp. 1–3
  • An adjustable routing threshold lets users move along the diversity-quality spectrum, while different routing strategies provide an additional control axis.The strategies use complementary signals and have different strengths across diversity dimensions.Source: §1 · p. 2
  • 21.3% improvement over the strongest inference-time baselines is achieved by BACO’s best router, consistently across tasks and evaluation metrics and with support from human evaluations.The paper reports this as a new state of the art for inference-time diversity-quality trade-offs.Source: §1 · p. 2

2. Preliminary

The preliminary section frames diversity and quality as a trade-off and motivates seeking a Pareto frontier rather than relying on either model alone. A Llama-3 pilot illustrates the models’ opposing strengths.

2. Preliminary

  • Diversity is measured over a group of outputs generated from the same prompt, while group-level quality is the average of per-output quality scores.Per-output quality may be assessed by a reward model or human evaluator.Source: §2 · p. 3
  • 3.15x more diversity for the base model and 5.95x higher quality for the aligned model illustrate the trade-off in a Llama-3 pilot on WildChat.The pilot measured diversity by semantic-equivalent output clusters and quality by average reward from another LLM.Source: Table 1 · p. 3, §2 · p. 3
  • The paper frames each method and configuration as a point in a two-dimensional diversity-quality space, with an ideal method approximating the Pareto frontier.On this frontier, diversity cannot improve without sacrificing quality, and vice versa.Source: §2 · pp. 3–4

3. BACO: Base-Aligned Collaboration for Diversity and Quality

BACO coordinates a base model and its aligned counterpart at the token level, routing generation toward diversity or quality as appropriate. Its routing strategies use uncertainty or predicted-token content, with a threshold to adjust the balance.

3. BACO: Base-Aligned Collaboration for Diversity and Quality

  • BACO pairs a base model as a source of diversity with its aligned counterpart as a source of quality, and a lightweight router chooses which model generates at each decoding step.The router acts as a gatekeeper, selecting the base when diversity is desired and the aligned model when quality is desired.Source: §3 · p. 4
  • Switching is restricted to word boundaries because a word may contain multiple tokens and the paired models may use different tokenizations.The paper points to Appendix B for full decoding pseudocode and further implementation details.Source: §3 · p. 4

3.1. Routing Strategy Design

  • Each routing strategy uses a threshold γ to adjust the balance: larger values bias toward the base model, while smaller values bias toward the aligned model.Strategies specify what information guides routing and which model to select from it.Source: §3.1 · p. 4
  • Logit-based strategies route using model uncertainty, while content-based strategies use predicted-token roles such as punctuation, formatting, or function words.Low maximum probability or high entropy can signal uncertainty; content-based strategies can also be used with black-box models because they do not require logits.Source: §3.1 · p. 4
  • Combining logit-based and content-based signals yields the strongest performance because the signals are complementary.BACO-P-FC, for example, prioritizes function-word routing before falling back to probability-based decisions.Source: §3.1 · pp. 4–5

4. Experimental Setup

BACO is evaluated on short- and long-form open-ended generation using diverse inference-time baselines and multiple diversity-quality measures. The evaluation aggregates trade-off curves across metric spaces, where -P-PUNC performs best overall among the reported BACO routers.

4. Experimental Setup

  • The evaluation spans NoveltyBench, WildChat, and Narrative-Discourse, covering instruction following, dialogue, and creative writing across short and long outputs.The datasets cover varying levels of task complexity.Source: §4 · p. 5
  • BACO is compared with single-model sampling, prompting, decoding, ensemble, and collaboration inference-time baselines.Some baselines, including diverse beam search, paraphrase prompting, and in-context resampling, require additional computation.Source: §4 · p. 5

4.1. Automatic Evaluations

  • Coverage summarizes the area under each method’s diversity-quality curve, while Dominance measures its contribution to the global Pareto frontier.Higher Coverage indicates stronger performance across a wider range of operating regions; higher Dominance indicates uniquely strong trade-offs.Source: §4.1 · p. 6
  • Eleven diversity metrics are paired with two quality metrics to form 11 × 2 spaces, with lexical and semantic results also analyzed separately.Overall results average across the spaces; quality is measured by aligned-model perplexity and reward-model scores.Source: §4.1 · p. 6
  • -P-PUNC achieves the best overall performance across NoveltyBench diversity-quality spaces, balancing lexical and semantic diversity more consistently than -RAND.-RAND’s gains concentrate on lexical metrics and fall sharply on semantic metrics.Source: §4.1 · p. 7

5. Results

Across datasets and evaluation spaces, BACO improves the diversity-quality trade-off, with especially strong semantic-diversity results. Router comparisons, model-pairing ablations, long-form evaluation, and human judgments further examine the framework’s performance.

5.1. Overall Performance

  • 0.142 Coverage improvement and 32.7% Dominance summarize BACO’s overall advantage across datasets and evaluation spaces.On NoveltyBench, Coverage improves by 0.274 and Dominance reaches 39.9%; semantic Dominance reaches 40.5%.Source: §5.1 · p. 7
  • BACO achieves higher diversity than the aligned baseline at matched quality or accuracy on IFEval and GSM8K.On GSM8K, naive temperature scaling degrades accuracy without meaningful diversity gains, while BACO improves diversity with accuracy maintained.Source: §5.1 · p. 7
  • On MT-Bench, BACO-P-PUNC reaches 0.681 Coverage and 72.8% Dominance, versus 0.320 and 27.2% for the aligned baseline.Source: §5.1 · p. 7

5.2. Router Performance Comparison

  • -P-PUNC has the strongest router controllability across lexical and semantic spaces; -RAND collapses on semantic metrics, while -P and -FC outperform it on semantic diversity.-H-PUNC and -P-FC capture additional Pareto-frontier regions, with 17.5% and 17.6% Dominance, respectively; BACO generally outperforms baselines across router choices.Source: §5.2 · p. 8

5.3. Base-Aligned vs. Aligned-Aligned Collaboration

  • On NoveltyBench, base-aligned collaboration significantly outperforms aligned-aligned collaboration in both Coverage and Dominance.The comparison supports the paper’s claim that the less-aligned base model contributes diversity not achieved by two aligned models.Source: §5.3 · p. 8

5.4. Long-Form Diversity Evaluation

  • At comparable quality, BACO achieves higher turning-point and arousal diversity than the aligned baseline in long-form creative writing.Source: §5.4 · p. 8

5.5. Human Evaluation

  • Human raters give BACO higher quality ratings than the aligned baseline on NoveltyBench and WildChat, and report significant diversity wins across format and content.Automatic quality scores are near-identical: 5.93 for aligned and 5.85 for BACO.Source: §5.5 · pp. 8–9
  • BACO is judged most creative in 79.6% of NoveltyBench prompts and 61.8% of WildChat prompts.The authors describe this creativity evaluation as preliminary and leave larger-scale investigation for future work.Source: §5.5 · p. 9

6. Analysis and Discussion

The analysis reports how base-model contribution and switching change during generation, discusses early stopping under diversity-aggressive routing, and outlines possible applications of controllable BACO.

6.1. Contribution Distribution and Switching Frequency

  • Base-model contribution and switching frequency are highest at generation start and decrease over time across all three tasks.The authors hypothesize that this reflects increasing predictive confidence as context grows, while noting the pattern may be less suitable for list-structured tasks.Source: §6.1 · p. 9

6.2. Failure Mode: Inherent Early Stopping

  • More diversity-aggressive routing increases the tendency to terminate early, which can truncate valid continuations or prevent incoherent repetition.Truncations are detectable by length, and the authors suggest restarting as a mitigation with marginal cost overhead.Source: §6.2 · p. 9

6.3. Future Work

  • The authors envision BACO as a controllable, inference-time-only preliminary step toward “breadth thinking” and as a possible plug-in for ideation agents.They also propose a user-facing slider for shifting along the diversity-quality spectrum.Source: §6.3 · p. 9

Impact Statement

BACO is intended for open-ended generation, not high-stakes or safety-critical domains, and should be deployed with standard content filtering or moderation because greater diversity may amplify harmful generations.

Impact Statement

  • BACO is designed for open-ended generation rather than high-stakes or safety-critical domains, and the authors recommend standard content filtering or moderation.They note that increased diversity might amplify harmful generations if safeguards are absent, a risk shared by most diversity-promoting technologies.Source: Impact Statement · p. 10
  • The human study used no sensitive personal data or personally identifiable information, and the datasets are public research benchmarks released under appropriate licenses.The authors also state that inference-time experiments without additional fine-tuning pose minimal memorization or privacy-leakage risks beyond those of the underlying models.Source: Impact Statement · p. 10
  • The authors describe BACO as complementary to safety alignment, not a replacement for safeguards.Source: Impact Statement · p. 10

Limitations

The paper identifies added inference cost and a focus on open-ended generation with canonical base and aligned checkpoints as limitations. Its related-work review contrasts diversity-promoting approaches and situates BACO among token-level collaboration methods.

Limitations

  • Collaboration adds FLOPs per token versus single-model inference, and the prototype is not yet speed-engineered.Speculative decoding, LoRA-based aligned models, and KV-cache sharing are proposed as mitigations.Source: Limitations · p. 16
  • The study focuses on inference-time improvements and open-ended tasks, leaving learned routers, RL-rollout diversity, agent and code generation, and longer horizons for future work.Source: Limitations · p. 16
  • The experiments mainly use canonical base and aligned checkpoints; intermediate alignment checkpoints may offer additional Pareto gains.Source: Limitations · p. 16

A. Related Work

  • Training-time diversity methods modify learning objectives but require retraining; inference-time alternatives can be costly or reduce quality, leaving stable trade-offs challenging.BACO is presented as an inference-time framework for more explicit control over that trade-off.Source: A. Related Work · p. 16
  • Prior multi-model collaboration targets quality or efficiency through weight-level, token-level, or multi-agent approaches; BACO advances token-level collaboration for the diversity-quality trade-off.Source: A. Related Work · pp. 16–17

B. BACO Framework Implementation and Router Details

BACO routes token generation between base and aligned models using ordered strategies, with implementation rules for token boundaries and termination. The framework incurs extra compute, while its routing family includes uncertainty- and content-based signals.

B. BACO Framework Implementation and Router Details

  • Tokenization alignment keeps tokens in a semantic unit with one model; default chat templates are used, with a plain shifting template for base models and tool calls and thinking disabled.Source: B. BACO Framework Implementation and… · p. 17
  • BACO evaluates routing strategies in order, selects the model chosen by the first strategy that fires, and guarantees termination with a final always-firing strategy.Single-strategy routers use one condition; combination routers use two or more, with complementary signals in practice.Source: B. BACO Framework Implementation and… · p. 17
  • At worst, BACO uses two forward passes per decoding step, about 2× FLOPs per token independent of group size; runtime is slower than single-model baselines.The prototype is comparable to or faster than several diversity baselines, including Diverse Beam Search and Logits Ensemble, but is not speed-engineered; listed optimizations may reduce cost.Source: B. BACO Framework Implementation and… · pp. 17–18, Table A6 · p. 18

B.1. Additional Rules in Router

  • The aligned model generates the first token; EOS is accepted only when it is also the aligned model’s top-1 prediction, limiting premature termination.The first-token choice is motivated by early-step influence and greater model disagreement; termination control also addresses possible degenerate continuations after switching.Source: §B.1 · p. 18

B.2. All Routing Strategies and Notations

  • Single-condition routers use base-model probability or entropy thresholds, aligned-model counterparts, probability or entropy ratios, content-word agreement, punctuation, or a judge model.The judge routes to base only when continuation can diverge and its sampled continuation is acceptable; random routing is also included.Source: §B.2 · pp. 18–19
  • Combined routers sequence content or punctuation rules with probability or entropy thresholds; aligned-logit metrics are less distinctive, while PUNC is lighter than FC and performs on-tier empirically.The judge strategy is a costlier extended comparison; FC adds part-of-speech parsing cost.Source: §B.2 · p. 19

C. Dataset Details

The evaluation draws on short, fine-grained diversity prompts, real open-ended conversations, and long-form creative writing. It samples groups of outputs and measures lexical and semantic diversity with multiple metrics.

C. Dataset Details

  • NoveltyBench evaluates multiple distinct, high-quality answers across randomness, underspecified facts, creative writing, and subjective queries, but its simple prompts often yield short outputs.The paper complements it with more complex datasets.Source: C. Dataset Details · p. 19
  • WildChat contributes real human–LLM prompts without fixed ground-truth answers, while Narrative-Discourse supports long-form fictional synopsis extensions and structural-diversity analysis.Source: C. Dataset Details · p. 20

D.1. Diversity Metrics

  • For each prompt, the study samples 10 outputs and evaluates group-level diversity with a broad set of automated metrics.Source: §D.1 · p. 20

D.1.1. LEXICAL METRICS

  • Lexical metrics include distinct-n, length- and vocabulary-normalized EAD-n, Self-BLEU, and Self-ROUGE-L; lower self-similarity scores indicate greater lexical diversity.Source: §D.1.1 · p. 20

D.1.2. SEMANTIC METRICS

  • Semantic metrics include embedding cosine dissimilarity, semantic entropy over entailment-based clusters, Vendi Score, NLI diversity, and NoveltyBench’s distinct functional-equivalence classes.Higher values indicate greater semantic diversity for Vendi Score and the distinct score; lower entailment indicates greater diversity for NLI diversity.Source: §D.1.2 · pp. 20–21

D.2. Aggregate Metrics over Quality–Diversity Spaces

The section defines quality–diversity spaces by metric pairs and compares methods using normalized Coverage and Dominance, then averages scores across spaces for metric-agnostic summaries.

D.2. Aggregate Metrics over Quality–Diversity Spaces

  • Each quality–diversity space pairs a higher-is-better quality metric with a diversity metric, and varying a method’s control parameter traces points in that space.Quality is higher-is-better; diversity may be higher- or lower-is-better depending on the metric.Source: §D.2 · p. 21
  • Coverage measures the area under a method’s normalized trade-off curve, indicating controllability and performance across parameter settings, but not whether it is ever optimal.Spaces are normalized to a unit square using base and aligned models at temperature 1.0 as reference points; points outside the feasible region are discarded.Source: §D.2 · pp. 21–22
  • Dominance uses the C-metric to measure a method’s share of the diversity- or quality-axis intervals on the global Pareto frontier, combining both axes by harmonic mean.Frontier responsibility intervals are split at midpoints between adjacent Pareto points.Source: §D.2 · p. 22
  • The method averages Coverage and Dominance over all spaces to produce metric-agnostic summaries.Source: §D.2 · p. 22

E. Experiment Setup

Experiments measure group-level diversity by generating ten outputs per prompt, using specified sampling settings unless a baseline requires a different decoding strategy.

E. Experiment Setup

  • For each prompt, the study generates a group of 10 outputs and ordinarily samples at temperature 1.0 with top-p 0.9 and no top-k truncation.Diverse Beam Search uses 20 beams, 10 beam groups, and diversity penalty 1.0; exceptions apply to baselines requiring alternative decoding.Source: E. Experiment Setup · p. 22

F. Validation on Verifiable Benchmarks

Across verifiable benchmarks and three generation settings, BACO reports higher diversity at matched quality or stronger diversity–quality trade-off results than baselines. The reported gains extend to longer dialogue and narrative generation.

F. Validation on Verifiable Benchmarks

  • On IFEval, BACO achieves consistently higher diversity than the aligned baseline at matched instruction-following accuracy.The benchmark uses automatically verifiable constraints and evaluates accuracy alongside diversity.Source: F. Validation on Verifiable Benchmarks · p. 23
  • On GSM8K, BACO maintains high accuracy, including 90% Pass@10, while achieving substantially higher diversity across operating points.Accuracy is measured with Acc and Pass@10; diversity is computed over sampled solutions.Source: F. Validation on Verifiable Benchmarks · p. 23

G.1. Instruction Following on NoveltyBench

  • BACO’s benchmark trends mirror those on open-ended tasks, and it also consistently outperforms baselines with the Olmo2 model family.Source: F. Validation on Verifiable Benchmarks · p. 23, §G.1 · p. 23
  • BACO improves Coverage by 0.274 overall (0.222 lexical, 0.291 semantic) on NoveltyBench and dominates 39.9% of the overall frontier, but does not lead lexical Dominance.It dominates 31.0% of the lexical and 48.8% of the semantic frontier.Source: §G.1 · p. 23

G.2. Dialogue on WildChat

  • On WildChat, BACO-P-PUNC improves Coverage by 29.8% over the aligned baseline and dominates 30.8% of the frontier, including 48.5% for semantic diversity.The reported results extend the comparison to longer, more conversational dialogue.Source: §G.2 · pp. 23–24

G.3. Creative Writing on Narrative-Discourse

  • On Narrative-Discourse, BACO achieves 13.5% higher Coverage and dominates 20.3% of the overall diversity–quality frontier.Source: §G.3 · p. 24

G.4. Extensional Tasks Results

The appendix compares BACO with the aligned baseline on IFEval and GSM8K, then describes discourse-level diversity measures for creative writing based on plot structure and emotional trajectories.

G.4. Extensional Tasks Results

H.1. Creative Writing Task Setup

  • Creative-writing diversity is measured through turning-point positions and emotional trajectories, capturing plot-structure and affective variation beyond surface-level metrics.Arousal diversity compares smoothed sentence-level arousal curves using KL divergence.Source: §H.1 · p. 25

H.2. Narrative-Discourse Evaluation Metrics

  • Turning-point diversity uses relative positions of five annotated plot inflections and pairwise distances across generated narratives.Source: §H.2 · p. 27
  • Arousal diversity is based on fixed-interval sentence samples scored by an LLM judge, smoothed by polynomial interpolation, and compared across narratives.Source: §H.2 · p. 27

I. Additional Analysis Material

The supplementary material documents additional analysis, human-evaluation procedures and examples, judge-prompt design, and two future-work directions.

I. Additional Analysis Material

  • Figures A9–A11 report per-token model contribution distributions and switching frequencies for Narrative-Discourse, WildChat, and NoveltyBench.The figures support the temporal-pattern observation discussed in §6.1.Source: §I · p. 27, Figure A9 · p. 29, Figure A10 · p. 29
  • Table A13 links one low-quality early stop among 10 samples to an incoherent continuation after a router switch; Table A14 says lower γ routes more to the aligned model and yields longer outputs.The A14 caption says higher γ increases routing to the base model, whose outputs are more likely to be incomplete and lower quality in this example.Source: Table A13 · p. 28, Table A14 · p. 28

I.1. Extended Future Work

  • Future work proposes speculative decoding to reduce inference overhead and exploring intermediate or partially aligned checkpoints for better diversity–quality operating points.The paper also suggests LoRA-tuned aligned models as a way to reduce deployment storage costs while preserving the contrast between checkpoints.Source: §I.1 · pp. 28–29

J.1. Setup

  • Human evaluation uses 20 prompts per dataset, three outputs per method, and four LLM-informed annotators on NoveltyBench and WildChat; Narrative-Discourse is excluded for output length.Prompts are stratified by average response length while maintaining comparable automatic diversity and quality scores.Source: §J.1 · pp. 29–30, Step 2: Compare Diversity (Per Column… · p. 30

Step 1: Evaluate Response Quality (Per Response)

  • Annotators rate each response from 1 to 5 on fluency, relevance, and substance, with brevity not penalized when appropriate and truncation not penalized when preceding text is meaningful.The rubric describes scores of 1, 3, and 5 as poor, adequate, and excellent, respectively.Source: Step 1: Evaluate Response Quality (Per… · p. 30

A. Overall Content Diversity

[Annotated as the more diverse group]

  • BACO demonstrates better human-evaluated quality, diversity, and creativity, with configurations matched on automatic quality metrics.The examples include individual output-quality scores and group-level diversity and creativity judgments.Source: [Annotated as the more diverse group] · pp. 33–34

Judge Model Prompt Design

  • The judge prompt asks whether multiple next-word continuations are valid, illustrating flexibility in open-ended tasks and constrained continuations in factual or context-bound examples.Its demonstrations contrast possible continuations for creative prompts with cases where a ground-truth answer or established context narrows the next word.Source: Judge Model Prompt Design · pp. 34–35

K. Qualitative Examples

The qualitative examples contrast repetitive or off-topic outputs with varied, prompt-responsive generations across job naming, dystopian food, story continuation, and Quasi’s trajectories from rebellion against the king to life as a pope. Their captions report BACo’s best diversity on NoveltyBench and best diversity–quality trade-offs on WildChat and on semantic entropy and perplexity.

K. Qualitative Examples

  • For the NoveltyBench prompt to name a reputed job, the base example repeats a disjointed story, while aligned examples name Data Scientist or Pilot and add career descriptions.The table caption says BACo shows the best diversity in semantic entropy; it does not provide the values.Source: K. Qualitative Examples · pp. 35–36, Table A18 · p. 37
  • The base model’s Soylent Bugs example is dominated by unrelated setting and instruction fragments, whereas the aligned example supplies a detailed food description and BACo offers a distinct snack-style presentation.The examples vary in framing, appearance, taste, and context rather than repeating a single description.Source: K. Qualitative Examples · pp. 38–40
  • On WildChat, BACo (Llama-3-8B and Llama-3-8B-Instruct) has semantic entropy 1.7247 and perplexity 8.5552; the aligned Llama-3-8B-Instruct model has 0.0000 and 1.8952, respectively.Source: K. Qualitative Examples · pp. 39–40, Table A19 · p. 41
  • For the Quasi story prompt, the aligned examples develop different complete endings, including a palace coup and a rebellion against King Guy.One ending makes Quasi join Catherine’s plan to overthrow the king; another has him plead to the Pope, flee with Catherine, and help defeat the king.Source: K. Qualitative Examples · p. 42
  • In one ending, Quasi rejects the king’s offer to become his personal torturer, rebels against his tyranny, and becomes a symbol of hope after Catherine takes the throne.Source: K. Qualitative Examples · p. 43
  • Quasi’s papal-drawing win brings fame, jealousy from Duchamp, and involvement in Catherine’s plan to unite the kingdoms before an attack disrupts the meeting.Source: K. Qualitative Examples · p. 44
  • In another trajectory, Catherine trains Quasi as a palace servant, he rises to Master of Ceremonies, and he succeeds Pope Cornelius as a peace-promoting pope.Source: K. Qualitative Examples · p. 44
  • Table A20 reports that BACO has the best diversity-quality trade-off on semantic entropy and perplexity among the base, aligned, and BACO outputs.The table caption says three outputs are randomly sampled from ten total.Source: Table A20 · p. 45
Loading 2511.05650v2…