Paper

{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2602.23730v1</id>\n    <title>Unlocking Cognitive Capabilities and Analyzing the Perception-Logic Trade-off</title>\n    <updated>2026-02-27T06:56:50Z</updated>\n    <link href='https://arxiv.org/abs/2602.23730v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2602.23730v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Recent advancements in Multimodal Large Language Models (MLLMs) pursue omni-perception capabilities, yet integrating robust sensory grounding with complex reasoning remains a challenge, particularly for underrepresented regions. In this report, we introduce the research preview of MERaLiON2-Omni (Alpha), a 10B-parameter multilingual omni-perception tailored for Southeast Asia (SEA). We present a progressive training pipeline that explicitly decouples and then integrates \"System 1\" (Perception) and \"System 2\" (Reasoning) capabilities. First, we establish a robust Perception Backbone by aligning region-specific audio-visual cues (e.g., Singlish code-switching, local cultural landmarks) with a multilingual LLM through orthogonal modality adaptation. Second, to inject cognitive capabilities without large-scale supervision, we propose a cost-effective Generate-Judge-Refine pipeline. By utilizing a Super-LLM to filter hallucinations and resolve conflicts via a consensus mechanism, we synthesize high-quality silver data that transfers textual Chain-of-Thought reasoning to multimodal scenarios.\n  Comprehensive evaluation on our newly introduced SEA-Omni Benchmark Suite reveals an Efficiency-Stability Paradox: while reasoning acts as a non-linear amplifier for abstract tasks (boosting mathematical and instruction-following performance significantly), it introduces instability in low-level sensory processing. Specifically, we identify Temporal Drift in long-context audio, where extended reasoning desynchronizes the model from acoustic timestamps, and Visual Over-interpretation, where logic overrides pixel-level reality. This report details the architecture, the data-efficient training recipe, and a diagnostic analysis of the trade-offs between robust perception and structured reasoning.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.AI'/>\n    <published>2026-02-27T06:56:50Z</published>\n    <arxiv:primary_category term='cs.AI'/>\n    <author>\n      <name>Longyin Zhang</name>\n    </author>\n    <author>\n      <name>Shuo Sun</name>\n    </author>\n    <author>\n      <name>Yingxu He</name>\n    </author>\n    <author>\n      <name>Won Cheng Yi Lewis</name>\n    </author>\n    <author>\n      <name>Muhammad Huzaifah Bin Md Shahrin</name>\n    </author>\n    <author>\n      <name>Hardik Bhupendra Sailor</name>\n    </author>\n    <author>\n      <name>Heng Meng Jeremy Wong</name>\n    </author>\n    <author>\n      <name>Tarun Kumar Vangani</name>\n    </author>\n    <author>\n      <name>Yi Ma</name>\n    </author>\n    <author>\n      <name>Qiongqiong Wang</name>\n    </author>\n    <author>\n      <name>Minh Duc Pham</name>\n    </author>\n    <author>\n      <name>Ridong Jiang</name>\n    </author>\n    <author>\n      <name>Jingtao Li</name>\n    </author>\n    <author>\n      <name>Jingyi Liao</name>\n    </author>\n    <author>\n      <name>Zhuohan Liu</name>\n    </author>\n    <author>\n      <name>Yanfeng Lu</name>\n    </author>\n    <author>\n      <name>Manas Gupta</name>\n    </author>\n    <author>\n      <name>Ai Ti Aw</name>\n    </author>\n  </entry>"
}