MoCA-Agent: A Market-of-Claims Code Agent for Financial and Numerical Reasoning

Financial and tabular question answering requires more than fluent reasoning: answers must be grounded in the exact facts, formulas, units, signs, and scales that support them. A single misread cell or incorrect operation can silently produce a plausible but wrong result. We introduce \textsc{MOCA-Agent}, a market-of-claims code agent that replaces free-form multi-agent debate with claim-level verification. The system decomposes each question into typed atomic claims, asks specialist trader agents to buy or sell those claims, clears their orders into confidence-weighted accept/reject decisions, and synthesizes an executable Python program from market-supported evidence. A code-aware verifier then checks the program for execution, structural consistency, and common financial reasoning errors, with at most one market-aware repair round. Across ten public benchmarks spanning financial numerical reasoning, general tabular reasoning, ESG question answering, and multimodal chart reasoning, \textsc{MOCA-Agent} achieves strong performance using a fixed Qwen3.6-27B backbone, including $78.3\%$ on FinQA, $76.0\%$ on FinanceMath, $71.2\%$ on MultiHiertt, $86.9\%$ on ESGenius, and $85.6\%$ average on FinChart-Bench. These results show that aggregating evidence at the level of atomic claims, rather than whole answers, improves robustness in high-stakes numerical reasoning.\footnote{The code and data are available: https://github.com/UBC-NLP/MoCA-Agent.

Argus-Retriever: Vision-LLM Late-Interaction Retrieval with Region-Aware Query-Conditioned MoE for Visual Document Retrieval

Late-interaction vision-language retrievers represent each document page as many visual token embeddings and score queries with MaxSim. In systems such as ColPali, ColQwen, ColNomic, and Nemotron ColEmbed, the document embeddings are produced without seeing the query, so the same page is represented identically for a table lookup, a chart question, and a layout-sensitive evidence request. We introduce \textbf{Argus}, a family of query-conditioned late-interaction retrievers built on Qwen3.5-VL. Argus adds a region-aware Mixture-of-Experts module: the query encoder produces both retrieval embeddings and a compact context vector, the document page is pooled into spatial regions, and a query-aware router selects latent experts per region before MaxSim. The output remains a multi-vector index compatible with ColPali-style retrieval, but the document representation is now dependent on the query (i.e., $\mathbf{D}(q)$). All Argus models use a 1024-dimensional retrieval head, compared with the 2560-dimensional and 4096-dimensional heads of recent state-of-the-art systems, and are trained on roughly 9\% of the available public supervision rather than the full pool. The 9B model reaches \textbf{92.67} NDCG@5 on ViDoRe V1 and \textbf{86.0} NDCG@5 on the combined V1+V2 leaderboard, the highest reported value for an open late-interaction model on the combined leaderboard. Wrapped in a Qwen3.6-27B agentic retrieval pipeline on ViDoRe V3, Argus-9B further improves its NDCG@10 from 60.28 to \textbf{64.80} over public tasks, showing that the same retriever serves both as a strong standalone system and as a search primitive for iterative LLM agents.

BracketRank: Large Language Model Document Ranking via Reasoning-based Competitive Elimination

Reasoning-intensive retrieval requires deep semantic inference beyond surface-level keyword matching, posing a challenge for current LLM-based rerankers limited by context constraints and order sensitivity. We propose \textbf{\BracketRank}, a framework that treats document reranking as a reasoning-driven competitive tournament. Our approach introduces three key innovations: (1) adaptive grouping based on model context limits, (2) reasoning-enhanced prompts that mandate step-by-step relevance explanations, and (3) a bracket-style elimination structure with winner and loser tracks. This design ensures robust document advancement while enabling parallel processing across competition stages. Evaluation on the BRIGHT reasoning benchmark shows that \BracketRank achieves \textbf{26.56 nDCG@10}, significantly outperforming state-of-the-art baselines including RankGPT-4 (17.0) and Rank-R1-14B (20.5). On TREC datasets, BracketRank achieves 77.90 nDCG@5 on DL 19 and 75.85 nDCG@5 on DL 20, exceeding all baselines, establishing that explicit reasoning within competitive elimination is a powerful paradigm for complex, multi-step retrieval tasks. https://github.com/DataScienceUIBK/BracketRank

MARVEL: Multimodal Adaptive Reasoning-intensiVe Expand-rerank and retrievaL

Multimodal retrieval over text corpora remains a fundamental challenge: the best vision-language encoder achieves only 27.6 nDCG@10 on MM-BRIGHT, a reasoning-intensive multimodal retrieval benchmark, underperforming strong text-only systems. We argue that effective multimodal retrieval requires three tightly integrated capabilities that existing approaches address only in isolation: expanding the query's latent intent, retrieving with a model trained for complex reasoning, and reranking via explicit step-by-step reasoning over candidates. We introduce \textbf{MARVEL} (\textbf{M}ultimodal \textbf{A}daptive \textbf{R}easoning-intensi\textbf{V}e \textbf{E}xpand-rerank and retrieva\textbf{L}), a unified pipeline that combines LLM-driven query expansion, \textbf{MARVEL-Retriever} -- a reasoning-enhanced dense retriever fine-tuned for complex multimodal queries -- and GPT-4o-based chain-of-thought reranking with optional multi-pass reciprocal rank fusion. Evaluated on MM-BRIGHT across 29 technical domains, MARVEL achieves \textbf{37.9} nDCG@10, surpassing the best multimodal encoder by \textbf{+10.3 points} and outperforming all single-stage baselines in 27 of 29 domains and matching or approaching the best baseline in the remaining two highly-specialized domains (Crypto, Quantum Computing), demonstrating that reasoning-intensive multimodal retrieval is best addressed through a unified expand-retrieve-rerank framework. https://github.com/mm-bright/multimodal-reasoning-retrieval

BRIDGE: Multimodal-to-Text Retrieval via Reinforcement-Learned Query Alignment

Multimodal retrieval systems struggle to resolve image-text queries against text-only corpora: the best vision-language encoder achieves only 27.6 nDCG@10 on MM-BRIGHT, underperforming strong text-only retrievers. We argue the bottleneck is not the retriever but the query -- raw multimodal queries entangle visual descriptions, conversational noise, and retrieval intent in ways that systematically degrade embedding similarity. We present \textbf{BRIDGE}, a two-component system that resolves this mismatch without multimodal encoders. \textbf{FORGE} (\textbf{F}ocused Retrieval Query Generato\textbf{r}) is a query alignment model trained via reinforcement learning, which distills noisy multimodal queries into compact, retrieval-optimized search strings. \textbf{LENS} (\textbf{L}anguage-\textbf{E}nhanced \textbf{N}eural \textbf{S}earch) is a reasoning-enhanced dense retriever fine-tuned on reasoning-intensive retrieval data to handle the intent-rich queries FORGE produces. Evaluated on MM-BRIGHT (2,803 queries, 29 domains), BRIDGE achieves \textbf{29.7} nDCG@10, surpassing all multimodal encoder baselines including Nomic-Vision (27.6). When FORGE is applied as a plug-and-play aligner on top of Nomic-Vision, the combined system reaches \textbf{33.3} nDCG@10 -- exceeding the best text-only retriever (32.2) -- demonstrating that \textit{query alignment} is the key bottleneck in multimodal-to-text retrieval. https://github.com/mm-bright/multimodal-reasoning-retrieval

HIVE: Query, Hypothesize, Verify An LLM Framework for Multimodal Reasoning-Intensive Retrieval

Multimodal retrieval models fail on reasoning-intensive queries where images (diagrams, charts, screenshots) must be deeply integrated with text to identify relevant documents -- the best multimodal model achieves only 27.6 nDCG@10 on MM-BRIGHT, underperforming even strong text-only retrievers (32.2). We introduce \textbf{HIVE} (\textbf{H}ypothesis-driven \textbf{I}terative \textbf{V}isual \textbf{E}vidence Retrieval), a plug-and-play framework that injects explicit visual-text reasoning into a retriever via LLMs. HIVE operates in four stages: (1) initial retrieval over the corpus, (2) LLM-based compensatory query synthesis that explicitly articulates visual and logical gaps observed in top-$k$ candidates, (3) secondary retrieval with the refined query, and (4) LLM verification and reranking over the union of candidates. Evaluated on the multimodal-to-text track of MM-BRIGHT (2,803 real-world queries across 29 technical domains), HIVE achieves a new state-of-the-art aggregated nDCG@10 of \textbf{41.7} -- a \textbf{+9.5} point gain over the best text-only model (DiVeR: 32.2) and \textbf{+14.1} over the best multimodal model (Nomic-Vision: 27.6), where our reasoning-enhanced base retriever contributes 33.2 and the HIVE framework adds a further \textbf{+8.5} points -- with particularly strong results in visually demanding domains (Gaming: 68.2, Chemistry: 42.5, Sustainability: 49.4). Compatible with both standard and reasoning-enhanced retrievers, HIVE demonstrates that LLM-mediated visual hypothesis generation and verification can substantially close the multimodal reasoning gap in retrieval. https://github.com/mm-bright/multimodal-reasoning-retrieval

Are LLM-Based Retrievers Worth Their Cost? An Empirical Study of Efficiency, Robustness, and Reasoning Overhead

Large language model retrievers improve performance on complex queries, but their practical value depends on efficiency, robustness, and reliable confidence signals in addition to accuracy. We reproduce a reasoning-intensive retrieval benchmark (BRIGHT) across 12 tasks and 14 retrievers, and extend evaluation with cold-start indexing cost, query latency distributions and throughput, corpus scaling, robustness to controlled query perturbations, and confidence use (AUROC) for predicting query success. We also quantify \emph{reasoning overhead} by comparing standard queries to five provided reasoning-augmented variants, measuring accuracy gains relative to added latency. We find that some reasoning-specialized retrievers achieve strong effectiveness while remaining competitive in throughput, whereas several large LLM-based bi-encoders incur substantial latency for modest gains. Reasoning augmentation incurs minimal latency for sub-1B encoders but exhibits diminishing returns for top retrievers and may reduce performance on formal math/code domains. Confidence calibration is consistently weak across model families, indicating that raw retrieval scores are unreliable for downstream routing without additional calibration. We release all code and artifacts for reproducibility.

MM-BRIGHT: A Multi-Task Multimodal Benchmark for Reasoning-Intensive Retrieval

Existing retrieval benchmarks primarily consist of text-based queries where keyword or semantic matching is usually sufficient. Many real-world queries contain multimodal elements, particularly, images such as diagrams, charts, and screenshots that require intensive reasoning to identify relevant documents. To address this gap, we introduce MM-BRIGHT, the first multimodal benchmark for reasoning-intensive retrieval. Our dataset consists of 2,803 real-world queries spanning 29 diverse technical domains, with four tasks of increasing complexity: text-to-text, multimodal-to-text, multimodal-to-image, and multimodal-to-multimodal retrieval. Extensive evaluation reveals that state-of-the-art models struggle across all tasks: BM25 achieves only 8.5 nDCG@10 on text-only retrieval, while the best multimodal model Nomic-Vision reaches just 27.6 nDCG@10 on multimodal-to-text retrieval actually underperforming the best text-only model (DiVeR: 32.2). These results highlight substantial headroom and position MM-BRIGHT as a testbed for next-generation retrieval models that better integrate visual reasoning. Our code and data are available at https://github.com/mm-bright/MM-BRIGHT. See also our official website: https://mm-bright.github.io/.

TEMPO: A Realistic Multi-Domain Benchmark for Temporal Reasoning-Intensive Retrieval

Existing temporal QA benchmarks focus on simple fact-seeking queries from news corpora, while reasoning-intensive retrieval benchmarks lack temporal grounding. However, real-world information needs often require reasoning about temporal evolution and synthesizing evidence across time periods. We introduce TEMPO, the first benchmark combining temporal reasoning with reasoning-intensive retrieval across 13 domains. TEMPO features: (1) 1,730 complex queries requiring deep temporal reasoning such as tracking changes, identifying trends, or comparing cross-period evidence; (2) step-wise retrieval planning with 3,976 decomposed steps and gold documents mapped to each step for multi-hop evaluation; and (3) novel temporal metrics including Temporal Coverage@k and Temporal Precision@k measuring whether results span required time periods. Evaluation of 12 retrieval systems reveals substantial challenges: the best model (DiVeR) achieves only 32.0 NDCG@10 and 71.4\% Temporal Coverage@10, demonstrating difficulty in retrieving temporally complete evidence. We believe TEMPO provides a challenging benchmark for improving temporal reasoning in retrieval and RAG systems. Our code and data are available at https://github.com/tempo-bench/Tempo. See also our official website: https://tempo-bench.github.io/.

DeAR: Dual-Stage Document Reranking with Reasoning Agents via LLM Distillation

Large Language Models (LLMs) have transformed listwise document reranking by enabling global reasoning over candidate sets, yet single models often struggle to balance fine-grained relevance scoring with holistic cross-document analysis. We propose \textbf{De}ep\textbf{A}gent\textbf{R}ank (\textbf{\DeAR}), an open-source framework that decouples these tasks through a dual-stage approach, achieving superior accuracy and interpretability. In \emph{Stage 1}, we distill token-level relevance signals from a frozen 13B LLaMA teacher into a compact \{3, 8\}B student model using a hybrid of cross-entropy, RankNet, and KL divergence losses, ensuring robust pointwise scoring. In \emph{Stage 2}, we attach a second LoRA adapter and fine-tune on 20K GPT-4o-generated chain-of-thought permutations, enabling listwise reasoning with natural-language justifications. Evaluated on TREC-DL19/20, eight BEIR datasets, and NovelEval-2306, \DeAR surpasses open-source baselines by +5.1 nDCG@5 on DL20 and achieves 90.97 nDCG@10 on NovelEval, outperforming GPT-4 by +3.09. Without fine-tuning on Wikipedia, DeAR also excels in open-domain QA, achieving 54.29 Top-1 accuracy on Natural Questions, surpassing baselines like MonoT5, UPR, and RankGPT. Ablations confirm that dual-loss distillation ensures stable calibration, making \DeAR a highly effective and interpretable solution for modern reranking systems.\footnote{Dataset and code available at https://github.com/DataScienceUIBK/DeAR-Reranking.}.