{"paper":{"title":"Symbiotic-MoE: Unlocking the Synergy between Generation and Understanding","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Symbiotic-MoE lets generative training improve rather than degrade understanding in multimodal models through shared experts and staged optimization.","cross_cats":["cs.CL","cs.LG"],"primary_cat":"cs.CV","authors_text":"Liefeng Bo, Miles Yang, Ping Tan, Xiangyue Liu, Zhao Zhong, Zijian Zhang","submitted_at":"2026-04-09T03:19:26Z","abstract_excerpt":"Empowering Large Multimodal Models (LMMs) with image generation often leads to catastrophic forgetting in understanding tasks due to severe gradient conflicts. While existing paradigms like Mixture-of-Transformers (MoT) mitigate this conflict through structural isolation, they fundamentally sever cross-modal synergy and suffer from capacity fragmentation. In this work, we present Symbiotic-MoE, a unified pre-training framework that resolves task interference within a native multimodal Mixture-of-Experts (MoE) Transformers architecture with zero-parameter overhead. We first identify that standa"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Symbiotic-MoE resolves task interference within a native multimodal Mixture-of-Experts (MoE) Transformers architecture with zero-parameter overhead... boosting inherent understanding with remarkable gains on MMLU and OCRBench.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That partitioning experts into task-specific groups with shared experts as a semantic bridge will allow generative signals to enrich understanding without routing collapse or negative interference, and that the progressive training will reliably convert early volatility into constructive feedback.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Symbiotic-MoE introduces modality-aware expert disentanglement and progressive training in a multimodal MoE to achieve synergistic generation and understanding without task interference or extra parameters.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Symbiotic-MoE lets generative training improve rather than degrade understanding in multimodal models through shared experts and staged optimization.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"4771df1dad5b86d026d520a4e9157da03f0d35744620d035958589ce1822c949"},"source":{"id":"2604.07753","kind":"arxiv","version":2},"verdict":{"id":"56e09401-c286-4092-b473-eff63ceea866","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T18:19:49.478109Z","strongest_claim":"Symbiotic-MoE resolves task interference within a native multimodal Mixture-of-Experts (MoE) Transformers architecture with zero-parameter overhead... boosting inherent understanding with remarkable gains on MMLU and OCRBench.","one_line_summary":"Symbiotic-MoE introduces modality-aware expert disentanglement and progressive training in a multimodal MoE to achieve synergistic generation and understanding without task interference or extra parameters.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That partitioning experts into task-specific groups with shared experts as a semantic bridge will allow generative signals to enrich understanding without routing collapse or negative interference, and that the progressive training will reliably convert early volatility into constructive feedback.","pith_extraction_headline":"Symbiotic-MoE lets generative training improve rather than degrade understanding in multimodal models through shared experts and staged optimization."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.07753/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"a6e46ac44a8bc3784c9c9fd795fab92330fbb639a194e45bba224de9ba0514cf"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}