1 00:00:03,030 --> 00:00:05,400 Evan: Welcome to Daily Paper Cast. 2 00:00:05,448 --> 00:00:13,128 Ashley: Today's paper is from the Hugging Face daily paper list of September 30, 2026, with 52 upvotes. 3 00:00:13,176 --> 00:00:27,236 Evan: It's titled 'VoxMem: Benchmarking Multimodal Memory in Large Audio Language Models', written by Yang Xiao and Vidhyasaharan Sethu, with corresponding author Eun-Jung Holden from the University of Melbourne. 4 00:00:27,288 --> 00:00:27,988 Ashley: Great. 5 00:00:27,988 --> 00:00:30,388 Let's dive into the Introduction section. 6 00:00:30,432 --> 00:00:37,962 Evan: Large audio language models, or LALMs, have rapidly advanced the field of natural spoken interaction. 7 00:00:37,962 --> 00:00:44,272 These models enable dialogue systems to manage extended, multi-session conversations more effectively. 8 00:00:44,358 --> 00:00:45,288 Ashley: Evan. 9 00:00:45,288 --> 00:00:51,028 And for these systems to be truly effective, they must possess long-term memory capabilities. 10 00:00:51,028 --> 00:00:59,088 This means being able to accumulate information across past interactions and use that information to ground present responses. 11 00:00:59,136 --> 00:01:12,576 Evan: While simply expanding the context window can allow models to process longer dialogue histories directly, it doesn't necessarily guarantee that the model can reliably retrieve and use the information contained within those histories. 12 00:01:12,624 --> 00:01:13,554 Ashley: That's correct. 13 00:01:13,554 --> 00:01:21,084 As interaction histories grow, the critical challenge becomes whether information can be accurately retrieved and used across sessions. 14 00:01:21,084 --> 00:01:26,704 Evaluating spoken conversational memory and guiding future memory systems is essential. 15 00:01:26,860 --> 00:01:32,720 Evan: Existing benchmarks fall short because they generally lack a principled taxonomy of audio memory. 16 00:01:32,720 --> 00:01:41,320 They typically cover only a narrow range of spoken-memory capabilities, focusing predominantly on linguistic content and selected acoustic cues. 17 00:01:41,376 --> 00:01:42,136 Ashley: Right. 18 00:01:42,136 --> 00:01:49,026 And this results in evaluations that mainly assess memory for what was said, rather than how it was conveyed via audio. 19 00:01:49,026 --> 00:01:57,316 Such an approach makes it challenging to understand whether failures arise from particular evidence types, memory operations, or their interaction. 20 00:01:57,360 --> 00:02:08,160 Evan: VoxMem addresses these gaps by introducing a benchmark for evaluating spoken conversational memory across two complementary axes: acoustic evidence and memory operation. 21 00:02:08,208 --> 00:02:25,848 Ashley: Specifically, VoxMem defines four types of acoustic evidence: speech semantics, which is what was said, speaker identity, which is who said it, paralinguistic cues, which is how it was said, and environmental sound, covering what was heard around the speaker. 22 00:02:25,936 --> 00:02:39,326 Evan: And to capture how this information must be used, VoxMem defines four memory operations: information extraction, multi-session reasoning, temporal evolution tracking, and answer refusal. 23 00:02:39,326 --> 00:02:50,456 This helps in evaluating whether models can retrieve a single fact, integrate evidence across sessions, track changes over time, or refuse to answer when there's insufficient evidence. 24 00:02:50,520 --> 00:02:51,560 Ashley: Exactly. 25 00:02:51,560 --> 00:02:58,210 VoxMem evaluates each of these operations and evidence types across multi-session conversational histories. 26 00:02:58,210 --> 00:03:11,080 The benchmark consists of 3,196 evaluation instances spread over 34,743 spoken sessions, amounting to 177 hours of audio content. 27 00:03:11,136 --> 00:03:23,416 Evan: To enable controlled evaluation as context length scales, VoxMem spans four context budgets: 8K, 16K, 32K, and 64K tokens. 28 00:03:23,416 --> 00:03:33,136 Each setup ensures the question, answer, and evidence remain identical across different lengths, so any performance drop can be attributed to the growing context. 29 00:03:33,192 --> 00:03:45,742 Ashley: Testing across these various budgets revealed a significant finding: among 15 evaluated LALMs, none exceeded 40% overall accuracy at the 32K token budget. 30 00:03:45,742 --> 00:03:54,412 Models were particularly good at retaining 'what was said', but struggled with 'who said it', 'how it was said', and 'what was audible'. 31 00:03:54,456 --> 00:03:56,186 Evan: That's a substantial gap. 32 00:03:56,186 --> 00:04:04,736 Moreover, accuracy declines as the history grows across all evidence types, and error modes differ qualitatively across categories. 33 00:04:04,736 --> 00:04:11,676 VoxMem essentially aims to provide a foundation for measuring and driving progress in spoken conversational memory. 34 00:04:11,736 --> 00:04:14,956 Ashley: This brings us to the end of the Introduction section. 35 00:04:16,262 --> 00:04:25,422 Evan: Ashley, can you explain how VoxMem categorizes and evaluates the multimodal memory in large audio language models? 36 00:04:25,466 --> 00:04:32,756 Ashley: The VoxMem benchmark establishes a two-dimensional framework to evaluate spoken conversational memory systematically. 37 00:04:32,756 --> 00:04:41,126 On one axis, we have the types of acoustic evidence that must be retained, and on the other, the types of memory operations applied to this evidence. 38 00:04:41,186 --> 00:04:41,866 Evan: Got it. 39 00:04:41,866 --> 00:04:45,506 And how did the authors define these acoustic evidence types? 40 00:04:45,554 --> 00:04:48,984 Ashley: The benchmark targets four types of acoustic evidence. 41 00:04:48,984 --> 00:04:56,384 Firstly, speech semantics which is information recoverable from transcripts such as facts, plans, and preferences. 42 00:04:56,384 --> 00:04:59,294 This serves as the baseline for semantic memory. 43 00:04:59,354 --> 00:05:01,254 Evan: And what about the other types? 44 00:05:01,298 --> 00:05:08,568 Ashley: Secondly, we have speaker identity which requires the model to associate information with the voice of the speaker who produced it. 45 00:05:08,568 --> 00:05:18,008 Thirdly, paralinguistic cues cover how something was said, including vocal states and delivery characteristics like hesitation, emphasis, and laughter. 46 00:05:18,008 --> 00:05:29,678 Finally, environmental sound captures non-speech sound events such as traffic, alarms, and household noises—elements which are entirely audio-native and absent from transcripts. 47 00:05:29,738 --> 00:05:31,528 Evan: That’s quite comprehensive. 48 00:05:31,528 --> 00:05:34,338 Now, how are the memory operations defined? 49 00:05:34,424 --> 00:05:37,884 Ashley: VoxMem evaluates four distinct memory operations. 50 00:05:37,884 --> 00:05:43,864 Information extraction, multi-session reasoning, temporal evolution tracking, and answer refusal. 51 00:05:43,864 --> 00:05:49,634 Information extraction tests if the model can retrieve data from a specific session using acoustic evidence. 52 00:05:49,634 --> 00:05:54,084 Multi-session reasoning checks if the model can integrate evidence across sessions. 53 00:05:54,084 --> 00:05:59,234 Temporal evolution tracking evaluates the model’s ability to follow changes across sessions. 54 00:05:59,234 --> 00:06:04,234 Answer refusal checks if the model can abstain from answering when there’s insufficient evidence. 55 00:06:04,298 --> 00:06:05,658 Evan: Interesting. 56 00:06:05,658 --> 00:06:11,178 And how did the authors ensure naturalism and controlled scenarios in their benchmark construction? 57 00:06:11,234 --> 00:06:17,744 Ashley: VoxMem was constructed meticulously through multiple stages to ensure both naturalism and control. 58 00:06:17,744 --> 00:06:27,004 Each item was planned with a structured specification detailing the required evidence and valid sessions, before rendering them into user-assistant dialogues. 59 00:06:27,004 --> 00:06:33,794 During dialogue writing, user turns were authored independently from assistant turns to prevent answer leakage. 60 00:06:33,842 --> 00:06:36,902 Evan: But how did they translate this into spoken audio? 61 00:06:36,992 --> 00:06:44,402 Ashley: All sessions were synthesized using Higgs-TTS-3 with fixed VCTK voices for speakers. 62 00:06:44,402 --> 00:06:53,262 Paralinguistic cues were introduced via style controls, and environmental sounds were mixed in using assets from ESC-50. 63 00:06:53,306 --> 00:06:57,546 Evan: So, both the dialogue text and audio were carefully crafted. 64 00:06:57,546 --> 00:07:00,226 What about ensuring the context length scales? 65 00:07:00,290 --> 00:07:10,250 Ashley: To manage context scales, VoxMem includes histories at four token budgets: 8K, 16K, 32K, and 64K. 66 00:07:10,250 --> 00:07:16,280 These histories grow progressively while preserving the content, context, and session structure. 67 00:07:16,280 --> 00:07:25,390 Each history length strictly extends the shorter ones to prevent changes in the question itself, ensuring we’re genuinely measuring the impact of context length. 68 00:07:25,442 --> 00:07:28,952 Evan: Right, consistency is key in such evaluations. 69 00:07:28,952 --> 00:07:33,362 Can you elaborate on the evaluation protocol adopted for testing these models? 70 00:07:33,410 --> 00:07:34,270 Ashley: Sure. 71 00:07:34,270 --> 00:07:41,920 The evaluation aimed to measure models’ accuracy across all memory operations and evidence types at each context length. 72 00:07:41,920 --> 00:07:45,720 Models were given the interaction history followed by a query. 73 00:07:45,720 --> 00:07:54,250 Each response was judged for correctness using a Gemini model to avoid biases, particularly as Gemini models were among the evaluated ones. 74 00:07:54,354 --> 00:07:58,254 Evan: And how did they address potential biases in judging the responses? 75 00:07:58,298 --> 00:08:11,118 Ashley: Responses were scored using a rule-based agreement test between Gemini and GPT-5.6-Luna models, achieving a Cohen’s Kappa agreement of 0.95—that’s quite high. 76 00:08:11,162 --> 00:08:12,492 Evan: That's impressive. 77 00:08:12,492 --> 00:08:16,022 Any specific metrics they focused on during their analysis? 78 00:08:16,082 --> 00:08:17,232 Ashley: Indeed. 79 00:08:17,232 --> 00:08:21,942 They measured raw accuracy and retention relative to the 8K baseline. 80 00:08:21,942 --> 00:08:29,842 Accuracy declines were evident as history grew, which reinforces the complexity of maintaining reliable memory over longer histories. 81 00:08:29,976 --> 00:08:38,466 Evan: So essentially, while models retain semantic information better, they struggle with audio-native information as the context scales. 82 00:08:38,522 --> 00:08:39,452 Ashley: Precisely. 83 00:08:39,452 --> 00:08:50,152 Errors from these models also varied by operation type; for instance, temporal evolution tracking mostly failed due to localization breakdowns before any reasoning occurred. 84 00:08:50,152 --> 00:08:58,042 They also analyzed distinct error modes across different evidence types, showing varied underlying reasons for memory failures. 85 00:08:58,106 --> 00:09:03,156 Evan: This really shows the depth of insight VoxMem offers for improving LALMs. 86 00:09:03,156 --> 00:09:05,166 We’ve covered a lot about the method. 87 00:09:05,210 --> 00:09:07,810 Ashley: Yes, that wraps up the Method section. 88 00:09:09,115 --> 00:09:14,325 Evan: Ashley, let’s move on to the Experimental Setup and Results section of VoxMem. 89 00:09:14,325 --> 00:09:18,315 How did the authors design their experiments to evaluate these models? 90 00:09:18,363 --> 00:09:21,203 Ashley: The experimental setup was quite comprehensive. 91 00:09:21,203 --> 00:09:28,623 The authors evaluated 15 large audio language models, including ten open-weight and five proprietary systems. 92 00:09:28,623 --> 00:09:37,533 They ensured that each model was tested at four history lengths: 8K, 16K, 32K, and 64K tokens. 93 00:09:37,533 --> 00:09:42,763 This provides a controlled scale to see how performance changes as context length grows. 94 00:09:42,879 --> 00:09:47,279 Evan: And each model was evaluated through their native audio interfaces, right? 95 00:09:47,331 --> 00:09:48,531 Ashley: Exactly. 96 00:09:48,531 --> 00:09:53,161 Each model received user turns as audio and assistant turns as text. 97 00:09:53,161 --> 00:10:02,071 The final part of the interaction included an audio query, and models had to generate a single response based on the entire conversation history. 98 00:10:02,115 --> 00:10:05,095 Evan: How were the responses judged for correctness? 99 00:10:05,139 --> 00:10:14,419 Ashley: Responses were scored using a combination of exact string matching and a large language model judge, specifically the Gemini-3.7-Flash. 100 00:10:14,419 --> 00:10:21,019 This judge marked responses as correct or incorrect based on predefined answer types and rules. 101 00:10:21,075 --> 00:10:22,935 Evan: That sounds meticulous. 102 00:10:22,935 --> 00:10:25,855 So what kind of accuracy figures did they find? 103 00:10:25,899 --> 00:10:31,709 Ashley: Overall, no model exceeded 40% accuracy at the 32K token budget. 104 00:10:31,709 --> 00:10:39,189 The proprietary models averaged about 33%, while the open-weight models averaged around 21.9%. 105 00:10:39,189 --> 00:10:46,459 Across all evaluated models, performance remained far from what we could call reliable spoken conversational memory. 106 00:10:46,515 --> 00:10:52,475 Evan: Were there specific types of information that were easier or harder for the models to remember? 107 00:10:52,539 --> 00:10:57,959 Ashley: Yes, there was a marked difference between lexical and non-lexical information. 108 00:10:57,959 --> 00:11:08,279 For speech semantics, which involves the content of what was said, the models were relatively better, achieving an average of 55.6% accuracy. 109 00:11:08,279 --> 00:11:17,979 In contrast, speaker identity, paralinguistic cues, and environmental sound had much lower accuracies, often below 20%. 110 00:11:18,087 --> 00:11:24,207 Evan: So the models are much better at remembering what was said rather than who said it or how it was said. 111 00:11:24,267 --> 00:11:25,187 Ashley: Exactly. 112 00:11:25,187 --> 00:11:30,677 This gap highlights the difficulty LALMs have with non-lexical acoustic information. 113 00:11:30,677 --> 00:11:34,397 Even within non-lexical categories, there was variance. 114 00:11:34,397 --> 00:11:41,947 For instance, models performed a bit better with speaker identity compared to paralinguistic cues and environmental sounds. 115 00:11:42,003 --> 00:11:43,083 Evan: Interesting. 116 00:11:43,083 --> 00:11:47,603 Were there any insights about how these models handle different memory operations? 117 00:11:47,667 --> 00:11:53,447 Ashley: Yes, the researchers found that memory operations also influenced performance significantly. 118 00:11:53,447 --> 00:12:04,987 Multi-session reasoning was relatively robust, especially for speech semantics and speaker identity, indicating that models could handle combining evidence across sessions better than other operations. 119 00:12:05,043 --> 00:12:08,843 Evan: What about temporal evolution tracking and answer refusal? 120 00:12:08,907 --> 00:12:14,677 Ashley: Temporal evolution tracking, especially for non-lexical evidence, showed quite poor performance. 121 00:12:14,677 --> 00:12:23,397 For example, tracking paralinguistic cues or environmental sounds across sessions had very low accuracy rates, often collapsing altogether. 122 00:12:23,397 --> 00:12:34,787 On the other hand, answer refusal accuracy was higher for paralinguistic and environmental questions, possibly reflecting that models often abstain when they can't process acoustic information effectively. 123 00:12:34,851 --> 00:12:39,371 Evan: It appears the complexity of these tasks really challenges current models. 124 00:12:39,371 --> 00:12:42,331 Were there any additional error analyses conducted? 125 00:12:42,387 --> 00:12:47,607 Ashley: The authors performed fine-grained error analysis to understand why models failed. 126 00:12:47,607 --> 00:12:58,177 They categorized errors into several types, such as evidence recognition failures, binding errors, operation execution failures, and unsupported answers. 127 00:12:58,177 --> 00:13:06,307 For instance, many errors in information extraction tasks were due to models failing to correctly associate facts with the speakers who said them. 128 00:13:06,393 --> 00:13:11,023 Evan: And was there a pattern in how the models' performance degraded with longer histories? 129 00:13:11,067 --> 00:13:16,667 Ashley: Yes, for all evidence types, models' accuracy declined as history length increased. 130 00:13:16,667 --> 00:13:25,927 Speech semantics showed the highest retention rates, maintaining about 70.3% of their 8K token accuracy even at 64K tokens. 131 00:13:25,927 --> 00:13:32,307 In contrast, speaker identity and environmental sound had the most significant drop-offs in performance. 132 00:13:32,355 --> 00:13:37,335 Evan: So increasing the context window isn't a silver bullet for better performance. 133 00:13:37,395 --> 00:13:46,145 Ashley: Correct, simply expanding the context window doesn't address the inherent difficulties LALMs face with non-lexical acoustic information. 134 00:13:46,145 --> 00:13:52,795 The results suggest that more sophisticated mechanisms are needed for models to effectively use longer histories. 135 00:13:52,851 --> 00:13:59,951 Evan: This really highlights the gap between current capabilities and the goal of reliable long-term conversational memory. 136 00:14:00,003 --> 00:14:00,873 Ashley: Indeed. 137 00:14:00,873 --> 00:14:09,503 The experiments underline the need for improvement in how models deal with non-lexical acoustic evidence and handle various memory operations. 138 00:14:09,555 --> 00:14:13,315 Evan: This brings us to the end of the Experiment and Results section. 139 00:14:14,653 --> 00:14:17,213 Evan: Let's delve into the Related Work section. 140 00:14:17,213 --> 00:14:25,573 Ashley, can you give us an overview of how VoxMem positions itself among existing memory benchmarks for spoken dialogues? 141 00:14:25,667 --> 00:14:26,937 Ashley: Of course, Evan. 142 00:14:26,937 --> 00:14:36,217 VoxMem builds on various dimensions of memory evaluation for spoken dialogues that previous works have touched upon but never fully covered comprehensively. 143 00:14:36,269 --> 00:14:37,339 Evan: Interesting. 144 00:14:37,339 --> 00:14:40,269 How does it differentiate from earlier benchmarks? 145 00:14:40,325 --> 00:14:53,425 Ashley: Well, earlier benchmarks like SpokenWOZ and ContextDialog focus mainly on the lexical content of what was said, often neglecting the richness of the audio signal that captures how it was spoken or who said it. 146 00:14:53,477 --> 00:14:59,997 Evan: So, these earlier benchmarks miss out on evaluating the full scope of spoken conversational memory? 147 00:15:00,053 --> 00:15:01,373 Ashley: Exactly. 148 00:15:01,373 --> 00:15:14,373 Others, like AudioMarathon and LongSpeech, primarily target single long-form recordings to assess understanding of extended audio, rather than retaining spoken memory across distinct conversational sessions. 149 00:15:14,429 --> 00:15:15,539 Evan: That makes sense. 150 00:15:15,539 --> 00:15:19,409 And how does VoxMem tackle the issue of cross-session memory? 151 00:15:19,469 --> 00:15:23,579 Ashley: Cross-session memory is a crucial area that VoxMem addresses. 152 00:15:23,579 --> 00:15:28,729 Real spoken interactions are rarely confined to single, unbroken recordings. 153 00:15:28,729 --> 00:15:41,809 Instead, they unfold over multiple sessions, often with varying topics and contexts, making it essential to evaluate models on their ability to retrieve and integrate information across these separated instances. 154 00:15:41,861 --> 00:15:47,681 Evan: So, VoxMem ensures that this multi-session aspect is a core part of its evaluation? 155 00:15:47,741 --> 00:15:48,901 Ashley: Precisely. 156 00:15:48,901 --> 00:16:01,401 VoxMem stands out as the first benchmark built from the ground up to evaluate multi-session spoken conversational memory, with a diverse set of topics and conditions that mimic real-world interactions. 157 00:16:01,445 --> 00:16:06,845 Evan: Can you elaborate on the memory operations and how previous benchmarks have dealt with them? 158 00:16:06,893 --> 00:16:11,023 Ashley: Memory operations are about how remembered information must be used. 159 00:16:11,023 --> 00:16:14,803 Existing benchmarks rarely go beyond basic retrieval tasks. 160 00:16:14,803 --> 00:16:32,933 For instance, most focus on whether a model can retrieve specific information from a past session, but they don't evaluate more complex operations like integrating evidence from multiple sessions, tracking how information evolves, or recognizing when there isn't enough information to provide a coherent answer. 161 00:16:33,041 --> 00:16:36,001 Evan: And how does VoxMem enhance this aspect? 162 00:16:36,053 --> 00:16:48,613 Ashley: VoxMem enhances this by defining four distinct memory operations, as we touched on earlier: information extraction, multi-session reasoning, temporal evolution tracking, and answer refusal. 163 00:16:48,613 --> 00:16:59,393 By covering these, VoxMem ensures that models are tested not just on retrieving information, but on performing a variety of tasks that reflect the true complexity of human memory. 164 00:16:59,513 --> 00:17:05,613 Evan: Did the authors mention any qualitative failures in previous benchmarks that VoxMem aims to address? 165 00:17:05,669 --> 00:17:14,249 Ashley: Yes, the authors pointed out that earlier benchmarks often don’t account for structured failures that can tell us a lot about where and why a model is going wrong. 166 00:17:14,249 --> 00:17:27,669 Without a framework that characterizes both the types of memory operations and the kinds of acoustic evidence to be remembered, it's challenging to pinpoint whether a failure is due to the type of evidence, the operation applied, or their interaction. 167 00:17:27,785 --> 00:17:33,245 Evan: So VoxMem's comprehensive taxonomy helps in analyzing these structured failures? 168 00:17:33,293 --> 00:17:34,293 Ashley: Exactly. 169 00:17:34,293 --> 00:17:43,963 The detailed taxonomy in VoxMem allows for fine-grained error analysis, enabling researchers to identify specific failure modes and their underlying causes. 170 00:17:43,963 --> 00:17:47,633 This helps in developing targeted improvements for LALMs. 171 00:17:47,723 --> 00:17:56,973 Evan: It sounds like VoxMem not only fills gaps in earlier research but also sets a new standard for comprehensive and detailed evaluation in this field. 172 00:17:57,029 --> 00:17:58,019 Ashley: That's right. 173 00:17:58,019 --> 00:18:10,429 By addressing the limitations of previous benchmarks and introducing a detailed taxonomy, VoxMem offers a more holistic and rigorous evaluation framework for spoken conversational memory models. 174 00:18:10,493 --> 00:18:14,013 Evan: This brings us to the end of the Related Work section. 175 00:18:15,330 --> 00:18:20,270 Evan: Alright, Ashley, let's recap the key contributions and takeaways from this paper. 176 00:18:20,334 --> 00:18:21,684 Ashley: Certainly, Evan. 177 00:18:21,684 --> 00:18:34,274 VoxMem introduces a principled taxonomy for evaluating spoken conversational memory, covering both the types of acoustic evidence that need to be retained and the memory operations applied to this evidence. 178 00:18:34,326 --> 00:18:47,586 Evan: It focuses on four types of acoustic evidence: speech semantics, speaker identity, paralinguistic cues, and environmental sounds, ensuring a holistic evaluation that goes beyond just textual content. 179 00:18:47,646 --> 00:18:58,826 Ashley: Yes, and crucially, it evaluates four distinct memory operations: information extraction, multi-session reasoning, temporal evolution tracking, and answer refusal. 180 00:18:58,826 --> 00:19:05,346 This allows for a comprehensive assessment of how well models perform various tasks related to spoken memory. 181 00:19:05,406 --> 00:19:15,906 Evan: The benchmark is extensive, comprising 3,196 evaluation instances over 34,743 spoken sessions. 182 00:19:15,906 --> 00:19:29,566 The design ensures consistent testing across different context lengths—8K, 16K, 32K, and 64K tokens—which helps in observing how performance scales with increasing history. 183 00:19:29,622 --> 00:19:45,482 Ashley: Findings showed that none of the 15 evaluated models exceeded 40% accuracy at 32K tokens, highlighting significant challenges in retaining and using non-lexical acoustic information, particularly as context length grows. 184 00:19:45,534 --> 00:19:57,074 Evan: The error analysis provided valuable insights into the varied failure modes across different evidence types and memory operations, emphasizing areas where LALMs need improvement. 185 00:19:57,126 --> 00:20:07,946 Ashley: Overall, VoxMem sets a new standard for benchmarking the full scope of spoken conversational memory, providing a solid foundation for future research and development in this field. 186 00:20:07,998 --> 00:20:10,278 Evan: That's a wrap on today’s episode. 187 00:20:10,278 --> 00:20:12,818 Thanks for tuning in to Daily Paper Cast. 188 00:20:12,870 --> 00:20:15,550 Ashley: We hope you found this discussion insightful. 189 00:20:15,550 --> 00:20:23,430 Be sure to join us next time as we dive into more papers from the world of AI, NLP, CV, and beyond. 190 00:20:23,558 --> 00:20:29,818 Evan: Until then, keep exploring, stay curious, and we'll see you in the next episode.