Language Model Ensembles

We describe our system for SemEval-2026 Task 5, which requires rating the plausibility of given word senses of homonyms in short stories on a 5-point Likert scale. Systems are evaluated by the unweighted average of accuracy (within one standard deviation of mean human judgments) and Spearman Rank Correlation. We explore three prompting strategies using multiple closed-source commercial LLMs: (i) a baseline zero-shot setup, (ii) Chain-of-Thought (CoT) style prompting with structured reasoning, and (iii) a comparative prompting strategy for evaluating candidate word senses simultaneously. Furthermore, to account for the substantial inter-annotator variation present in the gold labels, we propose an ensemble setup by averaging model predictions. Our best official system, comprising an ensemble of LLMs across all three prompting strategies, placed 4th on the competition leaderboard with 0.88 accuracy and 0.83 Spearman's rho (0.86 average). Post-competition experiments with additional models further improved this performance to 0.92 accuracy and 0.85 Spearman's rho (0.89 average). We find that comparative prompting consistently improved performance across model families, and model ensembling significantly enhanced alignment with mean human judgments, suggesting that LLM ensembles are especially well suited for subjective semantic evaluation tasks involving multiple annotators.

This paper describes our system for SemEval-2026 Task 6, which classifies clarity of responses in political interviews into three categories: Clear Reply, Ambivalent, and Clear Non-Reply. We propose a heterogeneous dual large language model (LLM) ensemble via self-consistency (SC) and weighted voting, and a novel post-hoc correction mechanism, Deliberative Complexity Gating (DCG). This mechanism uses cross-model behavioral signals and exploits the finding that an LLM response-length proxy correlates strongly with sample ambiguity. To further examine mechanisms for improving ambiguity detection, we evaluated multi-agent debate as an alternative strategy for increasing deliberative capacity. Unlike DCG, which adaptively gates reasoning using cross-model behavioral signals, debate increases agent count without increasing model diversity. Our solution achieved a Macro-F1 score of 0.85 on the evaluation set, securing 3rd place and tied with the second-best reported score.

We present our system for SemEval-2026 Task 3 on dimensional aspect-based sentiment regression. Our approach combines a hybrid RoBERTa encoder, which jointly predicts sentiment using regression and discretized classification heads, with large language models (LLMs) via prediction-level ensemble learning. The hybrid encoder improves prediction stability by combining continuous and discretized sentiment representations. We further explore in-context learning with LLMs and ridge-regression stacking to combine encoder and LLM predictions. Experimental results on the development set show that ensemble learning significantly improves performance over individual models, achieving substantial reductions in RMSE and improvements in correlation scores. Our findings demonstrate the complementary strengths of encoder-based and LLM-based approaches for dimensional sentiment analysis. Our development code and resources will be shared at https://github.com/aaronlifenghan/ABSentiment

Large-scale content analysis is increasingly limited by the absence of observable ground truth or gold-standard labels, as creating such benchmarks through extensive human coding becomes impractical for massive datasets due to high time, cost, and consistency challenges. To overcome this barrier, we introduce the AI-CROWD protocol, which approximates ground truth by leveraging the collective outputs of an ensemble of large language models (LLMs). Rather than asserting that the resulting labels are true ground truth, the protocol generates a consensus-based approximation derived from convergent and divergent inferences across multiple models. By aggregating outputs via majority voting and interrogating agreement/disagreement patterns with diagnostic metrics, AI-CROWD identifies high-confidence classifications while flagging potential ambiguity or model-specific biases.

This paper introduces a new methodology for using LLM-based systems for accurate and efficient semantic tagging of UN Security Council resolutions. The main goal is to leverage LLM performance variability to build ensemble systems for data cleaning and semantic tagging tasks. We introduce two evaluation metrics: Content Preservation Ratio (CPR) and Tag Well-Formedness (TWF), in order to avoid hallucinations and unnecessary additions or omissions to the input text beyond the task requirement. These metrics allow the selection of the best output from multiple runs of several GPT models. GPT-4.1 achieved the highest metrics for both tasks (Cleaning: CPR 84.9% - Semantic Tagging: CPR 99.99% and TWF 99.92%). In terms of cost, smaller models, such as GPT-4.1-mini, achieved comparable performance to the best model in each task at only 20% of the cost. These metrics ultimately allowed the ensemble to select the optimal output (both cleaned and tagged content) for all the LLM models involved, across multiple runs. With this ensemble design and the use of metrics, we create a reliable LLM system for performing semantic tagging on challenging texts.

Practitioners have access to an abundance of language models and prompting strategies for solving many language modeling tasks; yet prior work shows that modeling performance is highly sensitive to both choices. Classical machine learning ensembling techniques offer a principled approach: aggregate predictions from multiple sources to achieve better performance than any single one. However, applying ensembling to language models during decoding is challenging: naively aggregating next-token probabilities yields samples from a locally normalized, biased approximation of the generally intractable ensemble distribution over strings. In this work, we introduce a unified framework for composing $K$ language models into $f$-ensemble distributions for a wide range of functions $f\colon\mathbb{R}_{\geq 0}^{K}\to\mathbb{R}_{\geq 0}$. To sample from these distributions, we propose a byte-level sequential Monte Carlo (SMC) algorithm that operates in a shared character space, enabling ensembles of models with mismatching vocabularies and consistent sampling in the limit. We evaluate a family of $f$-ensembles across prompt and model combinations for various structured text generation tasks, highlighting the benefits of alternative aggregation strategies over traditional probability averaging, and showing that better posterior approximations can yield better ensemble performance.

Code comment classification is a critical task for automated software documentation and analysis. In the context of the NLBSE'26 Tool Competition, we present LoRA-MME, a Multi-Model Ensemble architecture utilizing Parameter-Efficient Fine-Tuning (PEFT). Our approach addresses the multi-label classification challenge across Java, Python, and Pharo by combining the strengths of four distinct transformer encoders: UniXcoder, CodeBERT, GraphCodeBERT, and CodeBERTa. By independently fine-tuning these models using Low-Rank Adaptation(LoRA) and aggregating their predictions via a learned weighted ensemble strategy, we maximize classification performance without the memory overhead of full model fine-tuning. Our tool achieved an F1 Weighted score of 0.7906 and a Macro F1 of 0.6867 on the test set. However, the computational cost of the ensemble resulted in a final submission score of 41.20%, highlighting the trade-off between semantic accuracy and inference efficiency.

While large language models (LLMs) fine-tuned with lightweight adapters achieve strong performance across diverse tasks, their performance on individual tasks depends on the fine-tuning strategy. Fusing independently trained models with different strengths has shown promise for multi-task learning through three main strategies: ensembling, which combines outputs from independent models; merging, which fuses model weights via parameter averaging; and routing, which integrates models in an input-dependent fashion. However, many design decisions in these approaches remain understudied, and the relative benefits of more sophisticated ensembling, merging and routing techniques are not fully understood. We empirically evaluate their trade-offs, addressing two key questions: What are the advantages of going beyond uniform ensembling or merging? And does the flexibility of routing justify its complexity? Our findings indicate that non-uniform ensembling and merging improve performance, but routing offers even greater gains. To mitigate the computational cost of routing, we analyze expert selection techniques, showing that clustering and greedy subset selection can maintain reasonable performance with minimal overhead. These insights advance our understanding of model fusion for multi-task learning.

Financial prediction from long documents involves significant challenges, as actionable signals are often sparse and obscured by noise, and the optimal LLM for generating embeddings varies across tasks and time periods. In this paper, we propose FinAnchor(Financial Anchored Representations), a lightweight framework that integrates embeddings from multiple LLMs without fine-tuning the underlying models. FinAnchor addresses the incompatibility of feature spaces by selecting an anchor embedding space and learning linear mappings to align representations from other models into this anchor. These aligned features are then aggregated to form a unified representation for downstream prediction. Across multiple financial NLP tasks, FinAnchor consistently outperforms strong single-model baselines and standard ensemble methods, demonstrating the effectiveness of anchoring heterogeneous representations for robust financial prediction.

Artificial intelligence built on large foundation models has transformed language understanding, computer vision, and reasoning, yet these systems remain isolated and cannot readily share their capabilities. Coordinating the complementary strengths of independently developed, black-box foundation models is essential for trustworthy intelligent systems, yet no established method exists. Here we show that such coordination can be achieved through a meta-ensemble framework termed StackingNet, which aggregates the output predictions of independent models at inference. StackingNet improves accuracy, reduces individual-model error and group-wise disparities, ranks model reliability, and identifies or prunes models that degrade performance, all without access to internal parameters or training data. Across language comprehension, visual attribute estimation, and academic paper rating, it consistently outperforms individual models and classic ensembles, with gains that persist when the base models are uniformly strong. These gains stem from variance reduction and consensus alignment among independent models rather than from any emergent group cognition, and they widen as the model pool grows more diverse. By turning model diversity from a source of inconsistency into a resource for cooperation, StackingNet offers a practical path toward coordinated artificial intelligence, where progress emerges not only from larger single models but from principled cooperation among many specialized ones.

LLM-as-a-judge ensembles are the standard paradigm for scalable evaluation, but their aggregation mechanisms suffer from a fundamental flaw: they implicitly assume that judges provide independent estimates of true quality. However, in practice, LLM judges exhibit correlated errors caused by shared latent confounders -- such as verbosity, stylistic preferences, or training artifacts -- causing standard aggregation rules like majority vote or averaging to provide little gain or even amplify systematic mistakes. To address this, we introduce CARE, a confounder-aware aggregation framework that explicitly models LLM judge scores as arising from both a latent true-quality signal and shared confounding factors. Rather than heuristically re-weighting judges, CARE separates quality from confounders without access to ground-truth labels. We provide theoretical guarantees for identifiability and finite-sample recovery under shared confounders, and we quantify the systematic bias incurred when aggregation models omit confounding latent factors. Across 12 public benchmarks spanning continuous scoring, binary classification, and pairwise preference settings, CARE improves aggregation accuracy, reducing error by up to 26.8%. Code is released in https://github.com/SprocketLab/CARE.

We study neural multi-label classification under severe label imbalance through sentence-level detection of the 19 refined Schwartz human values in 74k English news and manifesto sentences (ValueEval'24 corpus). Each sentence carries a roughly balanced moral-presence label and a 19-way value annotation. First, moral presence is learnable from single sentences: a DeBERTa-base classifier reaches positive-class $F_1 \approx 0.73$ at the default threshold, which calibration does not improve. Second, comparing direct multi-label detectors with presence-gated hierarchies under an 8 GB consumer-grade GPU budget, we find that gating does not improve over direct prediction, as gate recall becomes a bottleneck. Third, studying lightweight auxiliary signals and small ensembles, we isolate decision-threshold calibration as a decisive, often overlooked factor: a standard text-only baseline already matches the best official ValueEval'24 English run at the default threshold (macro-$F_1 = 0.282$ vs. $\approx 0.28$), and tuning the threshold on validation alone raises it to $0.315$, most of our overall gain. Lightweight features do not survive a paired per-seed test; a soft-voting ensemble reaches our best macro-$F_1 = 0.332$. Calibration is not architecture-specific: it reproduces on RoBERTa-base, where the gain is larger ($+0.043$). To our knowledge, this is the first systematic comparison of direct and presence-gated architectures, lightweight feature-augmented encoders, and instruction-tuned Large Language Models (LLMs) at sentence level; benchmarked 7-9B LLMs (zero-/few-shot and QLoRA) lag behind the supervised ensemble under the same budget. We provide empirical guidance for compute-efficient, value-aware NLP models.

Large language models (LLMs) exhibit complementary strengths arising from differences in pretraining data, model architectures, and decoding behaviors. Inference-time ensembling provides a practical way to combine these capabilities without retraining. However, existing ensemble approaches suffer from fundamental limitations. Most rely on fixed fusion granularity, which lacks the flexibility required for mid-generation adaptation and fails to adapt to different generation characteristics across tasks. To address these challenges, we propose AdaFuse, an adaptive ensemble decoding framework that dynamically selects semantically appropriate fusion units during generation. Rather than committing to a fixed granularity, AdaFuse adjusts fusion behavior on the fly based on the decoding context, with words serving as basic building blocks for alignment. To be specific, we introduce an uncertainty-based criterion to decide whether to apply ensembling at each decoding step. Under confident decoding states, the model continues generation directly. In less certain states, AdaFuse invokes a diversity-aware scaling strategy to explore alternative candidate continuations and inform ensemble decisions. This design establishes a synergistic interaction between adaptive ensembling and test-time scaling, where ensemble decisions guide targeted exploration, and the resulting diversity in turn strengthens ensemble quality. Experiments on open-domain question answering, arithmetic reasoning, and machine translation demonstrate that AdaFuse consistently outperforms strong ensemble baselines, achieving an average relative improvement of 6.88%. The code is available at https://github.com/CCM0111/AdaFuse.

The rapid advancement of large language models (LLMs) has created a diverse landscape of models, each excelling at different tasks. This diversity drives researchers to employ multiple LLMs in practice, leaving behind valuable multi-LLM log data. This naturally leads to the question of whether such logs can be fully leveraged to fuse LLMs' complementary capabilities. Although prior work has explored various strategies for integrating multiple LLMs, we argue that practical fusion must meet two essential requirements: (1) compatibility with real-world serving scenarios (e.g., local and API-based serving), and (2) flexibility to operate at different stages of the LLM pipeline to meet varied user needs (e.g., fine-tuning and inference stages). To this end, we introduce LLMFusionBench, a large-scale benchmark for LLM fusion that spans 14 tasks across five domains, with responses from 20 open-source LLMs (8B--671B) totaling 103M tokens. Building on LLMFusionBench, we propose FusionFactory, a systematic framework with three elaborated levels: (1) query-level fusion via tailored LLM routers, (2) thought-level fusion leveraging retrieved abstract reasoning templates, and (3) model-level fusion via distillation from top-ranked responses. Experiments show that FusionFactory consistently outperforms the best individual LLM across all 14 benchmarks, with the optimal fusion configuration varying across benchmarks, highlighting the promise of multi-LLM log data as a practical foundation for fusing diverse LLM capabilities.

Generative Pretrained Transformers (GPTs) are foundational Large Language Models (LLMs) for text generation. However, individual LLMs often produce inconsistent outputs and exhibit biases, limiting their representation of diverse language patterns. The closed-source nature of many powerful LLMs further restricts industry applications due to data privacy concerns. Inspired by successes in text generation, LLM ensemble techniques are now increasingly explored for code generation. This article reviews these emerging ensemble approaches to enhance understanding, encourage further research, and promote practical implementation in both text and code generation. We categorize LLM ensembles into seven main methods - weight merging, knowledge fusion, mixture-of-experts, reward ensemble, output ensemble, routing, and cascading - analyzing capabilities of those approaches. Our findings highlight key benefits such as improved diversity representation, enhanced output quality, and greater application flexibility. These insights aid model selection for real-world tasks and crucially, lay groundwork for extending ensemble strategies to multimodal LLMs.

The advancement of LLMs and their accessibility have triggered renewed interest in multi-agent reinforcement learning as robust and adaptive frameworks for dynamically changing environments. This paper introduces RL-Focal, a two-stage RL agent framework that routes and ensembles LLMs. First, we develop the Decider RL-agent, which learns to dynamically select an ensemble of small size ($m_i$) among $N$ LLMs ($m_i \ll N$) for incoming queries from a user-defined downstream task $i$, by maximizing both error-diversity and reasoning-performance of the selected ensemble through iterative updates of task-adaptive rewards and policy. Second, to enable effective fusion of dynamically selected LLMs, we develop the stage-2 Fusion RL-agent, which learns to resolve reasoning conflicts from different LLMs and dynamically adapt to different ensemble teams composed by the Decider Agent for different downstream tasks. {\em Third}, we introduce the focal diversity metric to better model the error correlations among multiple LLMs further improving the generalization performance of the Decider Agent, which actively prunes the ensemble combinations. By focal diversity, we enhance performance across tasks by effectively promoting reward-aware and policy-adaptive ensemble selection and inference fusion. Extensive evaluations on five benchmarks show that RL-Focal achieves the performance improvement of 8.48% with an ensemble of small size compared to the best individual LLM in a pool and offers stronger robustness. Code is available here.

Transformer-based Large Language Models (LLMs) traditionally rely on final-layer loss for finetuning and final-layer representations for predictions, potentially overlooking the predictive power embedded in late layers. Interpretability tools such as the logit lens show that late-layer representations already carry largely formed, task-relevant predictions; here we ask whether that observation can be turned into an actionable training signal. We find that focusing tuning effort on these layers can yield losses comparable to those of the final layer, with complementary test-time behaviour. Building on this, we introduce a tuning framework, Mixture-of-Depths Ensemble (MoDE), which treats the late layers as an ensemble that contributes to the final logits through learned routing weights. MoDE can be applied on top of any existing tuning method (e.g., LoRA) and, in our experiments, modestly improves reasoning performance at a small parameter overhead. We present MoDE as a mechanism study showing that late-layer logits can be made directly useful for tuning, and that they can substitute for substantially larger trainable modules with comparable performance.