{"paper":{"title":"HyLaR: Hybrid Latent Reasoning with Decoupled Policy Optimization","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Hybrid latent reasoning interleaves discrete text generation with continuous visual states and optimizes the combination through decoupled policy learning.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Hao Zhang, Jinwen Luo, Shi-Zhe Chen, Tao Cheng, Yixin Qin, Zheng Wei","submitted_at":"2026-04-22T08:22:23Z","abstract_excerpt":"Chain-of-Thought (CoT) reasoning significantly elevates the complex problem-solving capabilities of multimodal large language models (MLLMs). However, adapting CoT to vision typically discretizes signals to fit LLM inputs, causing early semantic collapse and discarding fine-grained details. While external tools can mitigate this, they introduce a rigid bottleneck, confining reasoning to predefined operations. Although recent latent reasoning paradigms internalize visual states to overcome these limitations, optimizing the resulting hybrid discrete-continuous action space remains challenging. I"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"HyLaR outperforms standard MLLMs and state-of-the-art latent reasoning approaches across fine-grained perception and general multimodal understanding benchmarks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The hybrid discrete-continuous action space can be effectively optimized via DePO with independent trust-region constraints and exact closed-form vMF KL regularizer without introducing instabilities or requiring extensive tuning.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"HyLaR with DePO enables effective RL in hybrid discrete-continuous spaces for multimodal models, outperforming prior MLLMs on perception and understanding benchmarks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Hybrid latent reasoning interleaves discrete text generation with continuous visual states and optimizes the combination through decoupled policy learning.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"7937486afadb009a9b8b6a15e42628e18f8badd3b1015be6769c4df06f16bacf"},"source":{"id":"2604.20328","kind":"arxiv","version":2},"verdict":{"id":"feef6044-d18c-4823-b458-e36c7fff8f9a","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T01:14:42.494900Z","strongest_claim":"HyLaR outperforms standard MLLMs and state-of-the-art latent reasoning approaches across fine-grained perception and general multimodal understanding benchmarks.","one_line_summary":"HyLaR with DePO enables effective RL in hybrid discrete-continuous spaces for multimodal models, outperforming prior MLLMs on perception and understanding benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The hybrid discrete-continuous action space can be effectively optimized via DePO with independent trust-region constraints and exact closed-form vMF KL regularizer without introducing instabilities or requiring extensive tuning.","pith_extraction_headline":"Hybrid latent reasoning interleaves discrete text generation with continuous visual states and optimizes the combination through decoupled policy learning."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.20328/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-21T14:43:16.450368Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-20T02:01:52.429681Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"ef30cdec3e955be5ea350d70b217a387a2e5dc6afad95f658985cdb89f63033c"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}