Agentic Choice Enhancement (ACE)

Updated 4 July 2026

Agentic Choice Enhancement (ACE) is a lightweight, post-generation answer reranker that refines video-language model outputs for deeper cognitive questions.
It operates in two stages: generating multiple candidate responses via beam search and using a compact selector (Llama-3.2) to choose the best answer, achieving up to 25% quality improvement.
ACE addresses gaps in depth, evidence, and coherence during System-2 reasoning in movies without modifying base model parameters.

Searching arXiv for the primary ACE formulation and closely related usages to ground the article. Agentic Choice Enhancement (ACE) is a lightweight, post-generation answer refinement module introduced in MovieCORE for video-LLMs (VLMs) confronted with harder, System-2-style movie reasoning questions. It is defined as a training-free, inference-time reranking plugin: the base VLM generates multiple candidate answers, and a compact LLM selects the best candidate rather than trusting the default decoding output. In MovieCORE, ACE is presented as a practical remedy for the observed weakness of standard VLMs on deeper cognitive movie questions, especially along the dimensions of depth, evidence, and coherence, and it is reported to improve answer quality by up to 25% relative on supervised open-source systems (Faure et al., 26 Aug 2025).

1. Problem setting and motivation

ACE arises from MovieCORE’s central finding that standard VLMs, including strong open-source and fully supervised models, still struggle with deeper cognitive movie questions. MovieCORE is explicitly designed to move beyond surface-level “what happened?” prompts and instead probe “why,” “how,” and “what evidence supports this?” style reasoning. The benchmark targets character motivations, causal chains, emotional interpretation, thematic understanding, and evidential grounding from the video. Within that setting, the weakest dimensions for standard systems are often depth and evidence, and open-source models remain below proprietary models even after full supervision (Faure et al., 26 Aug 2025).

The rationale for ACE is therefore not architectural replacement, but post-training remediation. The paper positions ACE as “a second pair of eyes”: a base VLM may already contain a better answer among its plausible outputs, yet its default decoding choice may not be the strongest answer for a reasoning-heavy task. ACE is introduced to expose that latent quality at inference time without modifying the underlying model parameters. This makes it distinct from training-time adaptation, parameter-efficient fine-tuning, and prompt-only optimization. In the MovieCORE formulation, the intervention is specifically targeted at answer quality for open-ended, reasoning-intensive movie question answering rather than simple recognition (Faure et al., 26 Aug 2025).

2. Mechanism and formal procedure

ACE treats generation as a two-stage process. First, the base VLM performs candidate generation by beam search. Second, a separate selector model scores and reranks the generated candidates. The MovieCORE paper gives the procedure explicitly as Algorithm 1:

$\text{Input: Video } V,\ \text{Question } Q,\ \text{Beam width } k=5$

$\text{Output: Best response } R^*$

The algorithmic steps are:

Generate candidate responses from the VLM:

$C \gets \text{VLM.generate}(V, Q, \text{beam\_width}=k)$

Score each candidate with Llama-3.2:

$S \gets \text{Llama-3.2.score}(C)$

Select the highest-scoring candidate:

$R^* \gets \arg\max_{c \in C} S(c)$

Return $R^*$ (Faure et al., 26 Aug 2025).

The reranker is the compact 1B-parameter Llama-3.2, chosen for efficiency. No task-specific supervision for the reranker is described; it is prompted without explicit evaluation guidelines and relies on its general notion of answer quality. Architecturally, ACE is a post-generation refinement plugin rather than a learned architectural change. It does not modify the encoder, decoder, multimodal fusion layers, or training objective of the base VLM. Its practical pipeline is simply to feed video $V$ and question $Q$ to a VLM, use beam search with width $k=5$ , obtain multiple candidate answers, ask Llama-3.2 to score the candidates, and output the candidate with maximal score (Faure et al., 26 Aug 2025).

A frequent misconception is to treat ACE as ordinary beam search. MovieCORE explicitly rejects that interpretation. ACE is not “take the top beam,” nor is it greedy search, sampling, or plain beam search. Its defining step is an independent selector model that judges candidate generations. For that reason, the paper describes ACE as inference-time candidate generation plus learned reranking, rather than as a decoding trick or a fine-tuning method (Faure et al., 26 Aug 2025).

3. Evaluation protocol on MovieCORE

ACE is evaluated within MovieCORE’s multidimensional LLM-assisted evaluation framework, not only with top-1 correctness. The paper argues that strict top-1 accuracy alone is inadequate for this benchmark because many valid answers are paraphrastic or partially open-ended, especially for System-2 questions. Accordingly, answers are scored on five dimensions, each on a 0–5 scale: Accuracy, Comprehensiveness, Depth, Evidence, and Coherence. GPT-4o-mini is used as the judge for these metrics (Faure et al., 26 Aug 2025).

The evaluation spans proprietary baselines, open-source zero-shot models, fully supervised open-source models, and supervised models augmented with ACE. The paper lists Gemini 2.5-flash, Gemini-1.5-pro, and GPT-4o as proprietary baselines. The open-source zero-shot set includes InstructBLIP, MA-LMM, HERMES, LongVU, mPlug-Owl3, Qwen2.5-VL, InternVL2, and InternVL2.5. Fully supervised versions of InstructBLIP, MA-LMM, and HERMES are then compared against those same supervised backbones with ACE added (Faure et al., 26 Aug 2025).

For broader comparison, the paper also reports BLEU-4, CIDEr, and METEOR. These standard metrics follow the same qualitative trend as the 0–5 rubric, but the paper stresses that n-gram-oriented measures are less suitable for System-2 reasoning. This framing matters because ACE is not primarily presented as an exact-match booster; it is presented as a quality-oriented refinement mechanism for richer, better-grounded answers (Faure et al., 26 Aug 2025).

4. Reported gains and empirical profile

The strongest reported result is an up to 25% relative improvement in answer quality. This comes from InstructBLIP under supervision, where the average score rises from 2.63 to 3.29, for an absolute gain of +0.66. MA-LMM rises from 2.79 to 3.35, an absolute gain of +0.56 and a relative improvement of about 20%. HERMES rises from 2.93 to 3.41, an absolute gain of +0.48 and a relative improvement of about 16% (Faure et al., 26 Aug 2025).

Model	Supervised avg	+ACE avg
InstructBLIP	2.63	3.29
MA-LMM	2.79	3.35
HERMES	2.93	3.41

The dimension-level pattern is central to the MovieCORE interpretation of ACE. For HERMES, ACE improves all five evaluated dimensions: Accuracy 3.52 → 3.81, Comprehensiveness 2.72 → 3.30, Depth 2.83 → 3.12, Evidence 2.98 → 3.38, Coherence 2.62 → 3.42, and Average 2.93 → 3.41. The paper emphasizes that ACE does not merely raise one metric; it consistently improves reasoning depth, evidence, and coherence, which are the hardest dimensions on MovieCORE (Faure et al., 26 Aug 2025).

The same qualitative pattern appears under traditional automatic metrics. For HERMES, BLEU-4 improves from 0.0308 to 0.0654, CIDEr from 0.1230 to 0.1622, and METEOR from 0.0983 to 0.2138 when ACE is added. MovieCORE does not claim that ACE beats the proprietary models overall, but it does state that ACE meaningfully narrows the gap for open-source supervised systems. The comparison is always against the same backbone without ACE, so the reported gains are attributed to the selection mechanism rather than to a different base model (Faure et al., 26 Aug 2025).

5. Ablations, limits, and failure modes

MovieCORE includes a beam-size ablation for HERMES + ACE. With beam width 3, the average score is 3.45; with beam width 5, 3.41; with beam width 7, 3.37. The paper concludes that performance is similar across beam widths 3, 5, and 7, with no clear winner. This suggests that the benefit comes more from the agentic selection step itself than from the exact beam hyperparameter (Faure et al., 26 Aug 2025).

The paper does not present a detailed ACE-specific failure taxonomy, but several limitations are explicitly implied. First, ACE depends on the base VLM: if all beam candidates are poor, selection cannot fully rescue the output. Second, the reranker is not task-trained, so Llama-3.2 may not align perfectly with MovieCORE’s nuanced reasoning criteria. Third, evaluation is LLM-assisted, so improvements may partly reflect the judge model’s preferences and inherit judge bias. Fourth, the broader dataset is only partially human-verified, and ACE is evaluated on a benchmark built from MovieChat-1k clips, which may limit genre diversity (Faure et al., 26 Aug 2025).

A more subtle limitation concerns the interpretation of the gains. The strongest reported improvements are mostly in answer-quality dimensions rather than guaranteed correctness in an exact factual sense. This fits the benchmark’s open-ended structure: ACE improves richness, grounding, and coherence, but it is not presented as solving deep movie reasoning completely. A plausible implication is that ACE is best understood as a quality-sensitive refinement mechanism whose effectiveness is bounded by the quality and diversity of the base model’s candidate set (Faure et al., 26 Aug 2025).

6. Position within the broader ACE literature

The acronym “ACE” is not stable across recent arXiv literature, and this has created a recurrent source of confusion. In MovieCORE, ACE denotes Agentic Choice Enhancement: a post-generation answer reranker for VLMs. By contrast, “Agentic Context Evolution” introduces a multi-agent retrieve-or-think framework in which a central orchestrator decides between RETRIEVE and THINK by majority voting over agents, with the goal of evolving context selectively in multi-hop QA (Chen et al., 13 Jan 2026). Although both methods are agentic and inference-time, they address different bottlenecks: answer reranking in video question answering versus adaptive context management in retrieval-augmented reasoning.

A second distinct usage is “Action and Control via Explanations,” a human-centered explainability paradigm for multimodal AI assistants in manufacturing. There, ACE refers to LLM-generated semantic-frame explanations that let end users correct and enrich multimodal system understanding across vision, speech, and document inputs (Watkins et al., 27 Feb 2025). A third is “Adaptive Context Elasticizer,” a plug-and-play context manager that stores both raw and abstract historical messages and adaptively assigns each step the elastic type Raw, Abs, or Drop across ReAct, DeepAgent, WebThinker, and MiroFlow (Liao et al., 30 Jun 2026). These works are adjacent in their concern with agentic control, but they are not the MovieCORE module.

Related lines of work further broaden the conceptual field. “Choice via AI” formalizes recommendation under possible menu misinterpretation, distinguishing AIC, RAIC, and GRAIC through conditions such as NSC, double monotonicity, and idempotence (Kops et al., 4 Feb 2026). “M $^3$ -ACE” extends ACE-style context engineering to multimodal math reasoning by rectifying visual evidence through multi-agent collaboration around visual evidence lists rather than by direct answer voting (Xie et al., 9 Mar 2026). These papers show that the phrase “agentic choice enhancement” can refer either narrowly to the MovieCORE reranking plugin or more loosely to a family of methods that improve decisions by intervening on selection, context, interpretation, or evidence. In current usage, however, the most direct and explicit instantiation of Agentic Choice Enhancement as a named module is the MovieCORE answer-refinement mechanism (Faure et al., 26 Aug 2025).