---

%PDF-1.4 %� 7 0 obj (��. Introduction) endobj 8 0 obj << /D (section.2) /S /GoTo >> endobj 11 0 obj (��. Related Work) endobj 12 0 obj << /D (section.3) /S /GoTo >> endobj 15 0 obj (��. Approach) endobj 16 0 obj << /D (subsection.3.1) /S /GoTo >> endobj 19 0 obj (��. Preliminaries) endobj 20 0 obj << /D (subsection.3.2) /S /GoTo >> endobj 23 0 obj (��. Response Format Prompting) endobj 24 0 obj << /D (subsection.3.3) /S /GoTo >> endobj 27 0 obj (��. Scaling the Data and Model) endobj 28 0 obj << /D (subsection.3.4) /S /GoTo >> endobj 31 0 obj (��. Scaling to Higher Resolutions) endobj 32 0 obj << /D (section.4) /S /GoTo >> endobj 35 0 obj (��. Empirical Evaluation) endobj 36 0 obj << /D (subsection.4.1) /S /GoTo >> endobj 39 0 obj (��. Benchmarks) endobj 40 0 obj << /D (subsection.4.2) /S /GoTo >> endobj 43 0 obj (��. Results) endobj 44 0 obj << /D (subsection.4.3) /S /GoTo >> endobj 47 0 obj (��. Emerging Properties) endobj 48 0 obj << /D (subsection.4.4) /S /GoTo >> endobj 51 0 obj (��. Ablation on LLM Choices) endobj 52 0 obj << /D (section.5) /S /GoTo >> endobj 55 0 obj (��. Open Problems in LMMs) endobj 56 0 obj << /D (subsection.5.1) /S /GoTo >> endobj 59 0 obj (��. Data Efficiency) endobj 60 0 obj << /D (subsection.5.2) /S /GoTo >> endobj 63 0 obj (��. Rethinking Hallucination in LMMs) endobj 64 0 obj << /D (subsection.5.3) /S /GoTo >> endobj 67 0 obj (��. Compositional Capabilities) endobj 68 0 obj << /D (section.6) /S /GoTo >> endobj 71 0 obj (��. Conclusion) endobj 72 0 obj << /D (appendix.A) /S /GoTo >> endobj 75 0 obj (��. Implementation Details) endobj 76 0 obj << /D (subsection.A.1) /S /GoTo >> endobj 79 0 obj (��. LLaVA-1.5-HD) endobj 80 0 obj << /D (subsubsection.A.1.1) /S /GoTo >> endobj 83 0 obj (��Preprocessing) endobj 84 0 obj << /D (subsubsection.A.1.2) /S /GoTo >> endobj 87 0 obj (��Training) endobj 88 0 obj << /D (subsection.A.2) /S /GoTo >> endobj 91 0 obj (��. Data) endobj 92 0 obj << /D (subsection.A.3) /S /GoTo >> endobj 95 0 obj (��. Hyperparameters) endobj 96 0 obj << /D (appendix.B) /S /GoTo >> endobj 99 0 obj (��. Qualitative Results) endobj 100 0 obj << /D (subsection.B.1) /S /GoTo >> endobj 103 0 obj (��. Response Format Prompts) endobj 104 0 obj << /D (subsection.B.2) /S /GoTo >> endobj 107 0 obj (��. Compositional Capabilities) endobj 108 0 obj << /D (appendix.C) /S /GoTo >> endobj 111 0 obj (��. Limitations) endobj 112 0 obj << /D [ 113 0 R /Fit ] /S /GoTo >> endobj 154 0 obj << /Filter /FlateDecode /Length 4652 >> stream x
��ے�F�����#� ����ޕG�R����Ą�4�nbE\�T[o�_fH"�c��j�P�����͞g���W������W���L�$�gO<�0��Y��a�峇��C�v�?�M�s��z5_�i|_u���՝������׹̓�;U����;N�c����ü��Ӯ�=�?>�U�3c�2M��,,2;[�<,��Q�=6���]srߥ�,�<2B}��I��8�-�G����<�-X�vs�_N�-���aZ����;��97��u͑�y�+�Qh ~@�7���Q�������>u}e�, �5�๱a����g�u"�0�ܒ���}���T�$ �C��8�'���$B.���kv����j�t� ��2Ma¬0� ��u^�����v���m�K���a��S�s�mw+�1�J4�����?^��l,�r`}<�?�~��T�XԀcV�ϛ��?=ʹ¦��^W@�m)_�qX$�n�(ɜ�~���Tˣ�1�ˌ���X�53Q�3M�f[��]u��,xV�/l�����lە���nW�e��ߢ4z��=�XW��y�:yP/��q���H�n_��fQX��'B��ij�r$�B��w�����;Bȳh��;6$I� yf�"�������h�&�Z���,B˿���n:�ؐE=uqF����5�Ee�,ؤ^��q@�CNvȌ����MQ,���}���:6K7�;�V¥���jvs[����

---