1<!DOCTYPE HTML> 2<html lang="en"><head><meta http-equiv="Content-Type" content="text/html; charset=UTF-8"> 3 4 <title>Anisha Gunjal</title> 5 6 <meta name="author" content="Anisha Gunjal"> 7 <meta name="viewport" content="width=device-width, initial-scale=1"> 8 9 <link rel="stylesheet" type="text/css" href="stylesheet.css"> 10 <link rel="icon" type="image/png" href="images/AnishaGunjal.jpg"> 11 <meta name="google-site-verification" content="luXvX-PJEB6t3Pd_FnX9sB528CNj_j1cZD68V-1IAF0" /> 12 <!-- Global Site Tag (gtag.js) - Google Analytics -->
vendor: 65 bytes, lines 12-13
12 13 <script async src="https://www.googletagmanager.com/gtag/js?id=
13UA-83979019-1
vendor: 14 bytes, lines 13-15
13"></script> 14 15
15<script> 16
vendor: 139 bytes, lines 16-19
16window.dataLayer = window.dataLayer || []; 17 function gtag(){dataLayer.push(arguments);} 18 gtag('js', new Date()); 19 gtag('config', '
19UA-83979019-1
vendor: 5 bytes, lines 19-20
19'); 20
20</script>
20 21</head> 22 23<body> 24 <table style="width:100%;max-width:800px;border:0px;border-spacing:0px;border-collapse:separate;margin-right:auto;margin-left:auto;"><tbody> 25 <tr style="padding:0px"> 26 <td style="padding:0px"> 27 <table style="width:100%;border:0px;border-spacing:0px;border-collapse:separate;margin-right:auto;margin-left:auto;"><tbody> 28 <tr style="padding:0px"> 29 <td style="padding:2.5%;width:63%;vertical-align:middle"> 30 <p style="text-align:center"> 31 <name>Anisha Gunjal</name> 32 </p> 33 <p> 34 I am a Machine Learning Research Engineer at <a href="https://scale.com/" target="_blank">Scale AI</a>, where I research on reasoning and alignment in large language models. My research also focuses on building robust datasets and evaluation benchmarks to promote trustworthiness and safety in LLMs. I recently completed my M.S. in Computer Science at <a href="https://www.utexas.edu/" target="_blank">UT Austin</a>, where I was advised by <a href="https://www.cs.utexas.edu/~gdurrett/" target="_blank">Prof. Greg Durrett</a> at the <a href="https://taur.cs.utexas.edu" target="_blank">TAUR Lab</a>. At UT, my research centered on fact verification and evaluation for language model outputs. 35</p> 36 37<p> 38 Prior to this, I was a Visiting Scholar at the <a href="https://www.clvrai.com/" target="_blank">Cognitive Learning for Vision and Robotics Lab (CLVR)</a> at the University of Southern California, advised by <a href="https://viterbi-web.usc.edu/~limjj/" target="_blank">Prof. Joseph Lim</a>. I also served as the <strong>Lead Machine Learning Engineer</strong> for Documents AI at <a href="https://hyperverge.co" target="_blank">HyperVerge, Inc.</a>, where I built scalable OCR and information extraction systems deployed across South and Southeast Asia. 39</p> 40 41<p> 42 Beyond the realms of research, I love traveling, salsa dancing, and capturing the <a href="https://www.instagram.com/theladoggie/?hl=en" target="_blank">adventures of my canine companions</a>! 43</p> 44</p> 45 46 <p style="text-align:center"> 47 <a href="mailto:[email protected]">Email</a>  /  48 <a href="data/AnishaGunjal.pdf" target="_blank">CV</a>  /  49 <a href="https://scholar.google.com/citations?user=KNsZ1u4AAAAJ&hl=en&oi=sra" target="_blank">Scholar</a>  /  50 <a href="https://twitter.com/anisha_gunjal" target="_blank">Twitter</a>  /  51 <a href="http://github.com/anisha2102" target="_blank">Github</a>  /  52 <a href="https://www.linkedin.com/in/anisha-gunjal/" target="_blank">LinkedIn</a>  /  53 <a href="https://medium.com/@anishagunjal7" target="_blank">Blog</a> 54 </p> 55 </td> 56 <td style="padding:2.5%;width:40%;max-width:40%"> 57 <a href="images/AnishaGunjal.jpg"><img style="width:100%;max-width:100%" alt="profile photo" src="images/AnishaGunjal.jpg" class="hoverZoomLink"></a> 58 </td> 59 </tr> 60 </tbody></table> 61 62 <table style="width:100%;border:0px;border-spacing:0px;border-collapse:separate;margin-right:auto;margin-left:auto;"><tbody> 63 64 <tr> 65 <td style="padding:20px;width:100%;vertical-align:middle"> 66 <heading>Updates</heading> 67 68 <table id="updates" style='padding:10px 20px 0px 0px'> 69 <tr> 70 <td>Jun 2024</td> 71 <td style="padding-left: 30px;">Started full-time as a Machine Learning Research Engineer at <a href="https://scale.com/" target="_blank">Scale AI</a></td> 72 </tr> 73 <tr> 74 <td>May 2024</td> 75 <td style="padding-left: 30px;">Graduated with an M.S. in Computer Science from <a href="https://www.cs.utexas.edu/" target="_blank">UT Austin</a></td> 76 </tr> 77 78 79 <tr> 80 <tr> 81 <td>Jan 2022</td> 82 <td style="padding-left: 30px;"><a href="https://clvrai.github.io/tarp/" target="_blank"> Task-Induced Representation Learning</a> accepted at <a href="https://iclr.cc/" target="_blank"> ICLR 2022</a></td> 83 </tr> 84 <tr> 85 <td>Dec 2021</td> 86 <td style="padding-left: 30px;">Poster Presentation at <a href="https://sites.google.com/view/deep-rl-workshop-neurips2021" target="_blank"> Deep RL Workshop, NeurIPS 2021</a></td> 87 </tr> 88 <tr> 89 <td>Aug 2021</td> 90 <td style="padding-left: 30px;">Participant at <a href="http://cvit.iiit.ac.in/summerschool2021/index.php" target="_blank"> 5th Summer School on Artificial Intelligence</a>
90, IIIT Hyderabad </td> 91 </tr> 92 <tr> 93 <td>Jun 2021</td> 94 <td style="padding-left: 30px;">Joined <a href="https://www.clvrai.com/" target="_blank">CLVR Lab</a> at USC as a Visiting Researcher</td> 95 </tr> 96 <tr> 97 <td>Sep 2020</td> 98 <td style="padding-left: 30px;">Published a blogpost <a href='https://medium.com/@anishagunjal7/the-eccv2020-experience-8bd1f068f673' target="_blank">The ECCV Experience</a></td> 99 </tr> 100 <tr> 101 <td>Jul 2020</td> 102 <td style="padding-left: 30px;">Reviewer at <a href="https://jupytercon.com/talk-poster-cfp/" target="_blank">JupyterCon2020</a></td> 103 </tr> 104 <tr> 105 <td>Jul 2020</td> 106 <td style="padding-left: 30px;">Published blogpost <a href="https://medium.com/@anishagunjal7/document-visual-question-answering-e6090f3bddee" target="_blank">Document Visual Question Answering</td> 107 </tr> 108 109 </table> 110 111 </td> 112 </tr> 113 </tbody> 114 </table> 115 116 117 118 <table style="width:100%;border:0px;border-spacing:0px;border-collapse:separate;margin-right:auto;margin-left:auto;"><tbody> 119 <tr> 120 <td style="padding:20px;width:100%;vertical-align:middle"> 121 <heading>Research</heading> 122 <p> 123 My research lies at the intersection of post-training, alignment, and trustworthiness in large language models, with a focus on human-grounded machine learning. I develop methods that enhance the reasoning capabilities and interpretability of LLMs through post-training techniques such as RLHF, rubric-guided reward modeling, and preference optimization. A core part of my work involves constructing high-quality evaluation benchmarks and datasets that reflect real-world complexity and human values, enabling more reliable and transparent AI systems. 124 125 </span> 126 </p> 127 </td> 128 </tr> 129 </tbody></table> 130 <table style="width:100%;border:0px;border-spacing:0px;border-collapse:separate;margin-right:auto;margin-left:auto;"><tbody> 131 132 133 <tr> 134 <td style="padding:20px;width:25%;vertical-align:middle"> 135 <div class="one"> 136 <img src='images/guide.png' width="160"> 137 </div> 138 </td> 139 <td style="padding:20px;width:75%;vertical-align:middle"> 140 <papertitle>Adaptive Guidance Accelerates Reinforcement Learning of Reasoning Models</papertitle> 141 <br> 142 V. Nath, E. Lau, <strong>Anisha Gunjal</strong>, M. Sharma, N. Barhate, S. Hendryx 143 <br> 144 <em>arXiv preprint, 2025 </em> 145 <br> 146 <a href="https://www.arxiv.org/abs/2506.13923" target="_blank">paper</a> 147 <p>Introduces "Guide", a new class of reinforcement learning algorithms that integrates adaptive guidance to significantly improve reasoning in LLMs.</p> 148 </td> 149</tr> 150 151<tr> 152 <td style="padding:20px;width:25%;vertical-align:middle"> 153 <div class="one"> 154 <img src='images/molecular.png' width="160"> 155 </div> 156 </td> 157 <td style="padding:20px;width:75%;vertical-align:middle"> 158 <papertitle>Molecular Facts: Desiderata for Decontextualization in LLM Fact Verification</papertitle> 159 <br> 160 <strong>Anisha Gunjal</strong>, <a href="https://www.cs.utexas.edu/~gdurrett/" target="_blank">Greg Durrett</a> 161 <br> 162 <em>EMNLP Findings 2024</em> 163 <br> 164 <a href="https://arxiv.org/abs/2406.20079" target="_blank">paper</a> 165 <p>Studies how context affects LLM fact verification and introduces the concept of âmolecular factsâ to isolate and evaluate factual reasoning. </p> 166 </td> 167</tr> 168 169 170 <tr> 171 <td style="padding:20px;width:25%;vertical-align:middle"> 172 <div class="one"> 173 174 <img src='images/mhaldetect.png' width="160"> 175 </div> 176 177 </td> 178 <td style="padding:20px;width:75%;vertical-align:middle"> 179 <papertitle>Detecting and Preventing Hallucinations in Large Vision Language Models</papertitle> 180 <br> 181 <strong>Anisha Gunjal</strong>, Jihan Yin, Erhan Bas</a> 182 <br> 183 <em>AAAI 2024</em> 184 <br> 185 <a href="https://arxiv.org/abs/2308.06394" target="_blank">paper</a> 186 <p>Reducing Hallucinations in LVLMs using a novel benchmark dataset M-HalDetect which is used to train fine-grained reward models 187 capable of detecting unfaithful text generations for a given image context.</p> 188 </td> 189 </tr> 190 191 <tr> 192 <td style="padding:20px;width:25%;vertical-align:middle"> 193 <div class="one"> 194 195 <img src='images/event.png' width="160"> 196 </div> 197 198 </td> 199 <td style="padding:20px;width:75%;vertical-align:middle"> 200 <papertitle>Drafting Event Schemas using Language Models</papertitle> 201 <br> 202 <strong>Anisha Gunjal</strong>, <a href="https://www.cs.utexas.edu/~gdurrett/" target="_blank">Greg Durrett</a></a> 203 <br> 204 <em>arxiv preprint 2023</em> 205 <br> 206 <a href="https://arxiv.org/abs/2305.14847" target="_blank">paper</a> 207 <p>Generating and automatically evaluating lightly structures complex event schemas using large language models.</p> 208 </td> 209 </tr> 210 211 <tr> 212 <td style="padding:20px;width:25%;vertical-align:middle"> 213 <div class="one"> 214 215 <img src='images/tarp.gif' width="160"> 216 </div> 217 218 </td> 219 <td style="padding:20px;width:75%;vertical-align:middle"> 220 <papertitle>Task-Induced Representation Learning</papertitle> 221 <br> 222 <a href="https://junjungoal.github.io/" target="_blank">Jun Yamada</a>, 223 <a href="https://kpertsch.github.io/" target="_blank">Karl Pertsch</a>, 224 <strong>Anisha Gunjal</strong>, 225 <a href="https://viterbi-web.usc.edu/~limjj/" target="_blank">Joseph Lim</a> 226 <br> 227 <em>ICLR 2022</em> 228 <br> 229 <a href="https://clvrai.github.io/tarp/" target="_blank">project page</a> / 230 <a href="https://openreview.net/forum?id=OzyXtIZAzFv" target="_blank">paper</a> 231 <p>Investigation of using task information for learning representations for RL in visually complex scenes.</p> 232 </td> 233 </tr> 234 235 <tr> 236 <td style="padding:20px;width:25%;vertical-align:middle"> 237 <div class="one"> 238 239 <img src='images/docvqa.png' width="160"> 240 </div> 241 242 </td> 243 <td style="padding:20px;width:75%;vertical-align:middle"> 244 <papertitle>Document Visual Question Answering</papertitle> 245 <br> 246 <strong>Anisha Gunjal</strong>, 247 <a href="https://vipulgupta1011.github.io/" target="_blank">Vipul Gupta</a>, 248 <a href="https://sites.google.com/view/moinakb" target="_blank">Moinak Bhattacharya</a>, 249 Digvijay Singh 250 <br> 251 <em>CVPR 2020</em>, Leaderboard Rank: 4 252 <br> 253 <a href="https://github.com/anisha2102/docvqa" target="_blank">
253code</a> 254 / 255 <a href="https://medium.com/@anishagunjal7/document-visual-question-answering-e6090f3bddee" target="_blank">blog</a> 256 / 257 <a href="https://docvqa.org/" target="_blank">workshop</a> 258 <p>Joint modeling of text and layout information using transformers for Visual Question Answering on unstructured documents.</p> 259 </td> 260 </tr> 261 262 <tr> 263 <td style="padding:20px;width:25%;vertical-align:middle"> 264 <div class="one"> 265 266 <img src='images/dr.png' width="160"> 267 </div> 268 269 </td> 270 <td style="padding:20px;width:75%;vertical-align:middle"> 271 <papertitle>Diabetic Retinopathy Grading using Deep Siamese Network</papertitle> 272 <br> 273 <strong>Anisha Gunjal</strong> 274 <br> 275 <em>ICML 2018</em>, Poster Presentation 276 <br> 277 <a href="https://drive.google.com/file/d/1CpgQKTze-X4Wgd0OCKEv4KOhscbaW-J1/view" target="_blank">paper</a> 278 / 279 <a href="https://sites.google.com/view/wcb2018/home?authuser=0" target="_blank">workshop</a> 280 <p>Contrastive Learning on retinal images for determining the stage of Diabetic Retinopathy disease progression on a small sized medical image dataset.</p> 281 </td> 282 </tr> 283 </tbody></table> 284 285 <table style="width:100%;border:0px;border-spacing:0px;border-collapse:separate;margin-right:auto;margin-left:auto;"><tbody> 286 <tr> 287 <td style="padding:20px;width:100%;vertical-align:middle"> 288 <heading>Blogposts</heading> 289 290 <table id="blogposts" style='padding:10px 20px 0px 0px'> 291 <tr> 292 <td><li> <a href="https://medium.com/@anishagunjal7/document-visual-question-answering-e6090f3bddee" target="_blank">Document Visual Question Answering</a> </td> 293 </tr> 294 <tr> 295 <td><li><a href="https://medium.com/@anishagunjal7/the-eccv2020-experience-8bd1f068f673" target="_blank">The ECCV Experience</a> </td> 296 </tr> 297 298 </table> 299 300 </td> 301 </tr> 302 </tbody> 303 </table> 304 305 <table style="width:100%;border:0px;border-spacing:0px;border-collapse:separate;margin-right:auto;margin-left:auto;"><tbody> 306 <tr> 307 <td style="padding:20px;width:100%;vertical-align:middle"> 308 <heading>Listicles</heading> 309 310 <table id="listicles" style='padding:10px 20px 0px 0px'> 311 <tr> 312 <td><li><a href="https://github.com/anisha2102/awesome-neuro-symbolic-ai" target="_blank">Awesome Resource List for NeuroSymbolic Visual Reasoning</a> </td> 313 </tr> 314 315 316 </table> 317 318 </td> 319 </tr> 320 </tbody> 321 </table> 322 323 <table style="width:100%;border:0px;border-spacing:0px;border-collapse:separate;margin-right:auto;margin-left:auto;"><tbody> 324 <tr> 325 <td style="padding:0px"> 326 <br> 327 <p style="text-align:right;font-size:small;"> 328 Clone <a style="font-size: 14px" href="https://github.com/jonbarron/jonbarron_website">this</a> to overcome good website design FOMO! 329 </td> 330 </tr> 331 </tbody></table> 332 </td> 333 </tr> 334 </table> 335</body> 336 337</html>
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.