1 00:00:03,000 --> 00:00:05,640 Evan: Welcome to Daily Paper Cast. 2 00:00:05,688 --> 00:00:13,248 Ashley: Today's paper comes from the Hugging Face daily paper list of October 28, 2026, with 32 upvotes. 3 00:00:13,296 --> 00:00:20,476 Evan: The title of the paper is: Can Test-Time Memory Evolution Outperform Test-Time Training for Scientific Discovery? 4 00:00:20,520 --> 00:00:28,450 Ashley: The first two authors are Yepeng Liu and Shengmai Chen, and the corresponding author is Yepeng Liu from UC Santa Barbara. 5 00:00:28,512 --> 00:00:31,212 Evan: Let's dive straight into the Introduction section. 6 00:00:31,272 --> 00:00:39,972 Ashley: Scientific discovery often involves proposing hypotheses, evaluating them, and using the resulting evidence to guide further investigation. 7 00:00:39,972 --> 00:00:50,862 Advances in reasoning, coding, and agentic tool-use capabilities of large language models (LLMs) have enabled systems that automate part of this discovery process. 8 00:00:50,928 --> 00:00:52,678 Evan: That's really interesting. 9 00:00:52,678 --> 00:00:56,428 So, how do these systems function in the realm of scientific discovery? 10 00:00:56,472 --> 00:01:06,142 Ashley: In the iterative discovery loop, a failed attempt can reveal an implementation constraint, while a successful modification can suggest an idea worth applying elsewhere. 11 00:01:06,142 --> 00:01:12,702 A key design question is how to learn from this accumulated experience to improve subsequent search efficiently. 12 00:01:12,808 --> 00:01:20,558 Evan: Recent frameworks like Test-Time Training, or TTT, seem to provide solutions to this challenge, right? 13 00:01:20,616 --> 00:01:21,726 Ashley: Exactly. 14 00:01:21,726 --> 00:01:34,896 Recent test-time training frameworks, such as TTT-Discover and ThetaEvolve, enable open-weight LLMs to outperform inference-only systems based on frontier models in scientific discovery tasks. 15 00:01:34,896 --> 00:01:40,586 Specifically, these frameworks combine iterative search with test-time reinforcement learning. 16 00:01:40,632 --> 00:01:50,132 Evan: But I've heard that TTT introduces additional computational cost for online weight updates during the discovery process. 17 00:01:50,132 --> 00:01:52,262 How does that affect scalability? 18 00:01:52,320 --> 00:01:53,470 Ashley: You're right. 19 00:01:53,470 --> 00:02:01,790 TTT requires online model weight adaptation throughout the discovery process, which is computationally expensive and time-consuming. 20 00:02:01,790 --> 00:02:07,480 This makes it challenging to scale TTT to larger models for more complex discovery tasks. 21 00:02:07,566 --> 00:02:10,376 Evan: And this is where the paper proposes a new method? 22 00:02:10,440 --> 00:02:11,660 Ashley: Yes, indeed. 23 00:02:11,660 --> 00:02:21,130 The paper introduces Test-Time Memory Evolution, or TTME, a plug-and-play test-time adaptation method as an alternative to TTT. 24 00:02:21,130 --> 00:02:29,010 TTME continually updates the external memory by distilling experience to guide subsequent discovery without updating model weights. 25 00:02:29,064 --> 00:02:30,494 Evan: Sounds promising. 26 00:02:30,494 --> 00:02:33,664 How do they compare TTME with TTT? 27 00:02:33,720 --> 00:02:43,040 Ashley: To have a clean comparison, they evaluate TTME and TTT under the same discovery framework, replacing only the test-time adaptation method. 28 00:02:43,040 --> 00:02:52,240 Their evaluation shows that TTME outperforms TTT on all five mathematical tasks while significantly reducing computational cost. 29 00:02:52,296 --> 00:02:55,556 Evan: What kind of improvements are we talking about here? 30 00:02:55,638 --> 00:03:07,068 Ashley: TTME reduces computational cost by 82.8 to 84.6 percent in FLOPs and 47.7 to 58.4 percent in runtime. 31 00:03:07,068 --> 00:03:16,038 These results establish TTME as not only effective but also computationally efficient compared to TTT in the evaluated discovery settings. 32 00:03:16,104 --> 00:03:17,094 Evan: Great. 33 00:03:17,094 --> 00:03:19,764 What are the core contributions of this paper? 34 00:03:19,824 --> 00:03:27,234 Ashley: First, they investigate whether test-time memory evolution can outperform test-time weight updates in scientific discovery. 35 00:03:27,234 --> 00:03:38,864 Second, they introduce a plug-and-play self-evolving memory method that continually updates external memory from evaluated search experience to guide subsequent discovery with a fixed-weight model. 36 00:03:38,864 --> 00:03:45,164 Third, their experiments show TTME outperforms TTT across all tested tasks. 37 00:03:45,216 --> 00:03:52,596 Evan: So, TTME seems like a significant step forward in making scientific discovery more efficient. 38 00:03:52,596 --> 00:03:55,766 This concludes our coverage of the Introduction section. 39 00:03:57,086 --> 00:04:01,166 Evan: Alright, let's dive into the Method section of the paper. 40 00:04:01,226 --> 00:04:10,146 Ashley: The authors propose Test-Time Memory Evolution, or TTME, as an efficient test-time adaptation method for scientific discovery. 41 00:04:10,146 --> 00:04:18,786 Unlike Test-Time Training, or TTT, which requires online weight updates, TTME focuses on updating the external memory. 42 00:04:18,786 --> 00:04:27,626 This approach leverages the foundational knowledge, reasoning, and coding capabilities of large language models without altering the model's internal weights. 43 00:04:27,734 --> 00:04:28,784 Evan: Interesting. 44 00:04:28,784 --> 00:04:32,024 So, how exactly does TTME work? 45 00:04:32,090 --> 00:04:42,270 Ashley: TTME divides search experience into two components: practical lessons for implementing candidate solutions and algorithmic ideas for improving task scores. 46 00:04:42,270 --> 00:04:49,400 Consequently, TTME maintains two complementary memory modules: Guideline Memory and Strategy Memory. 47 00:04:49,536 --> 00:04:52,866 Evan: Could you explain these components a bit more? 48 00:04:52,922 --> 00:04:54,052 Ashley: Of course. 49 00:04:54,052 --> 00:05:01,442 Guideline Memory distills execution and evaluation feedback into shared operational guidelines for reliable implementation. 50 00:05:01,442 --> 00:05:06,312 It provides concrete actions and mechanisms based on past successes and failures. 51 00:05:06,432 --> 00:05:13,312 Evan: So, should I picture these guidelines as a reference manual for the LLM? 52 00:05:13,370 --> 00:05:14,240 Ashley: Exactly. 53 00:05:14,240 --> 00:05:22,630 For instance, a guideline could suggest avoiding certain argument types in functions or preferring specific algorithms under certain conditions. 54 00:05:22,630 --> 00:05:27,650 This memory is updated every few search steps by summarizing collected outcomes. 55 00:05:27,758 --> 00:05:29,648 Evan: And what about Strategy Memory? 56 00:05:29,714 --> 00:05:36,754 Ashley: Strategy Memory, on the other hand, preserves algorithmic ideas from modifications that have improved task scores. 57 00:05:36,754 --> 00:05:42,824 It organizes these ideas into reusable strategies that can combine and adapt for future searches. 58 00:05:42,824 --> 00:05:50,274 These strategies are updated concurrent with Guideline Memory, based on analyzed improvements observed in evaluated cases. 59 00:05:50,330 --> 00:05:56,170 Evan: How is the experience stored in these memories actually used during the search? 60 00:05:56,264 --> 00:06:01,934 Ashley: During candidate generation, Guideline Memory provides shared guidance for implementation. 61 00:06:01,934 --> 00:06:08,134 Strategy Memory selects up to N distinct strategies based on their potential to improve the parent program. 62 00:06:08,134 --> 00:06:13,454 Selected strategies are then adapted into parent-specific modification instructions. 63 00:06:13,514 --> 00:06:15,714 Evan: Can you give us an example of this process? 64 00:06:15,770 --> 00:06:16,510 Ashley: Sure. 65 00:06:16,510 --> 00:06:22,180 Let's say a parent program was modified successfully to produce a valid result with a higher score. 66 00:06:22,180 --> 00:06:29,570 The elements of this change, such as a new algorithm or a different combination approach, are documented in Strategy Memory. 67 00:06:29,570 --> 00:06:40,850 When a similar task comes up, TTME uses this stored knowledge to inform and guide the generation of new candidate solutions by adapting these strategies to the current context. 68 00:06:40,898 --> 00:06:42,688 Evan: Interesting mechanism. 69 00:06:42,688 --> 00:06:49,898 So within this framework, how are the cases where the model struggles to find effective solutions handled? 70 00:06:49,946 --> 00:06:50,986 Ashley: Good question. 71 00:06:50,986 --> 00:06:58,516 When no suitable strategy is selected, the parent context and existing Strategy Memory are used to propose new ideas. 72 00:06:58,516 --> 00:07:06,026 This approach ensures the framework isn't stuck repeating unsuccessful modifications and always seeks innovative solutions. 73 00:07:06,164 --> 00:07:12,784 Evan: What concrete steps are involved in updating the memory and generating new candidates? 74 00:07:12,872 --> 00:07:16,682 Ashley: Every few search steps, TTME updates the memories. 75 00:07:16,682 --> 00:07:28,192 For Guideline Memory, outcomes across parents are summarized, new lessons are distilled, and these guidelines are merged with existing entries, removing stale entries and enforcing capacity limits. 76 00:07:28,192 --> 00:07:37,462 For Strategy Memory, evaluated improvement cases are selected, notes extracted, grouped into strategy families, and entries updated correspondingly. 77 00:07:37,514 --> 00:07:39,614 Evan: And during the generation process? 78 00:07:39,704 --> 00:07:44,654 Ashley: Generation involves forming modification instructions guided by Strategy Memory. 79 00:07:44,654 --> 00:07:56,534 Direct Generation uses shared context from Guideline Memory to generate candidates, while Strategy-Guided Generation utilizes specific modification instructions to produce parent-specific solutions. 80 00:07:56,624 --> 00:08:01,274 Evan: Does this method provide any notable computational benefits? 81 00:08:01,352 --> 00:08:02,232 Ashley: Evan. 82 00:08:02,232 --> 00:08:13,392 TTME reduces running time by 47.7 to 58.4 percent and computational costs by 82.8 to 84.6 percent. 83 00:08:13,392 --> 00:08:19,912 This efficiency arises from avoiding online model weight updates and utilizing external memory efficiently. 84 00:08:19,970 --> 00:08:21,590 Evan: That sounds impressive. 85 00:08:21,590 --> 00:08:25,110 How are TTME's effectiveness and efficiency evaluated? 86 00:08:25,154 --> 00:08:37,124 Ashley: The authors evaluated TTME under two TTT frameworks, across five mathematical construction tasks and three models spanning 8 billion and 120 billion parameters. 87 00:08:37,124 --> 00:08:45,364 They preserved each framework's experimental configuration, comparing results generated by TTME to those reported by TTT. 88 00:08:45,410 --> 00:08:47,400 Evan: And what did the evaluations show? 89 00:08:47,450 --> 00:08:55,600 Ashley: TTME not only outperformed TTT in all evaluated tasks but did so with significantly lower computational demands. 90 00:08:55,600 --> 00:09:02,510 The detailed comparison demonstrated the efficiency of TTME, making it a promising alternative to TTT. 91 00:09:02,570 --> 00:09:05,420 Evan: That concludes our discussion of the Method section. 92 00:09:05,420 --> 00:09:13,290 Next up, we'll explore the experiments and results reported in the paper to understand the impacts and implications of TTME further. 93 00:09:14,547 --> 00:09:22,867 Evan: Moving on to the Experiment and Results section, let's delve into how the authors evaluated Test-Time Memory Evolution. 94 00:09:22,923 --> 00:09:33,103 Ashley: The authors conducted a thorough evaluation across two distinct frameworks, TTT-Discover and ThetaEvolve, using five mathematical construction tasks. 95 00:09:33,103 --> 00:09:38,533 They aimed to assess the effectiveness and efficiency of TTME compared to TTT. 96 00:09:38,665 --> 00:09:42,035 Evan: What models did they use for these experiments? 97 00:09:42,099 --> 00:09:53,329 Ashley: For their experiments, they employed three models: Qwen3-8B, gpt-oss-120b, and Distill-Qwen3-8B. 98 00:09:53,329 --> 00:09:58,539 These models span parameters ranging from 8 billion to 120 billion. 99 00:09:58,587 --> 00:10:01,677 Evan: And what tasks were used to measure the performance? 100 00:10:01,731 --> 00:10:13,351 Ashley: The mathematical tasks evaluated were the Erdős minimum overlap problem, the first, second, and third autocorrelation inequalities, and the maximal determinant problem also known as Hadamard. 101 00:10:13,351 --> 00:10:18,531 These tasks were chosen for their complexity and the challenge they pose to large language models. 102 00:10:18,579 --> 00:10:23,639 Evan: So how exactly were TTME and TTT compared? 103 00:10:23,691 --> 00:10:32,941 Ashley: They preserved each framework’s original experimental configuration, replacing only the test-time adaptation method — TTT with TTME. 104 00:10:32,941 --> 00:10:38,871 This allowed them to conduct a clean comparison by isolating the impact of the adaptation method itself. 105 00:10:38,981 --> 00:10:42,021 Evan: Okay, now let’s look at the results. 106 00:10:42,021 --> 00:10:45,331 How did TTME perform compared to TTT? 107 00:10:45,387 --> 00:10:50,077 Ashley: TTME outperformed TTT across all evaluated tasks and models. 108 00:10:50,077 --> 00:10:56,637 In each case, TTME achieved better final scores while significantly reducing computational costs. 109 00:10:56,771 --> 00:11:01,501 Evan: Can you share some of the specific metrics or figures they reported? 110 00:11:01,563 --> 00:11:20,143 Ashley: For instance, under the TTT-Discover framework with the Qwen3-8B model, TTME reduced runtime by 47.7 to 58.4 percent and computational cost by 82.8 to 84.6 percent in FLOPs compared to TTT. 111 00:11:20,143 --> 00:11:30,963 The first autocorrelation inequality task showed a runtime reduction from 11.76 hours with TTT to just 5.36 hours with TTME. 112 00:11:31,011 --> 00:11:34,531 Evan: That's quite a substantial improvement. 113 00:11:34,531 --> 00:11:38,881 How did TTME perform in terms of the task scores themselves? 114 00:11:38,931 --> 00:11:42,041 Ashley: TTME consistently achieved better scores. 115 00:11:42,041 --> 00:11:57,281 For example, in the first autocorrelation inequality task, TTME attained a score of 1.504329 compared to TTT's 1.50525 with the Qwen3-8B model. 116 00:11:57,281 --> 00:12:08,161 Similarly, for the second autocorrelation inequality, TTME scored 0.949248 against TTT's 0.9472. 117 00:12:08,211 --> 00:12:09,631 Evan: That's impressive. 118 00:12:09,631 --> 00:12:14,281 Did the paper discuss any specific benefits in task performance other than scores? 119 00:12:14,331 --> 00:12:15,261 Ashley: Indeed. 120 00:12:15,261 --> 00:12:23,631 The paper highlighted how TTME improved the mean program validity rates over TTT and search without TTT across the tasks. 121 00:12:23,631 --> 00:12:34,581 For example, in the AC2 task, TTME achieved a mean validity rate of 95.42 percent compared to TTT's 78.02 percent. 122 00:12:34,635 --> 00:12:42,345 Evan: It sounds like TTME not only reduces computational cost but also enhances performance robustness. 123 00:12:42,345 --> 00:12:44,655 Are there any ablation studies mentioned? 124 00:12:44,715 --> 00:12:54,035 Ashley: Yes, the authors conducted ablation studies to further analyze the contributions of the two memory components, Guideline Memory and Strategy Memory. 125 00:12:54,035 --> 00:13:04,945 The results revealed that removing either component worsened performance, demonstrating the benefits of combining both shared operational lessons and reusable algorithmic strategies. 126 00:13:04,995 --> 00:13:09,245 Evan: How about combining TTME with TTT? 127 00:13:09,245 --> 00:13:12,495 Did they explore this possibility? 128 00:13:12,555 --> 00:13:21,165 Ashley: They did, but interestingly, TTME alone achieved the best scores, and adding TTT actually worsened the results. 129 00:13:21,165 --> 00:13:28,355 This suggests that TTME is most effective as an alternative to TTT rather than a complementary method. 130 00:13:28,419 --> 00:13:30,149 Evan: That's quite insightful. 131 00:13:30,149 --> 00:13:35,079 Were there any additional observations or analyses regarding TTME's performance? 132 00:13:35,139 --> 00:13:45,139 Ashley: The paper noted that TTME selects combinations of existing ideas more frequently for further search and more often selects combinations that improve task scores. 133 00:13:45,139 --> 00:13:52,659 This indicated that TTME leverages external memory to carry productive combinations into subsequent exploration effectively. 134 00:13:52,767 --> 00:14:00,307 Evan: It seems clear that TTME offers both computational efficiency and enhanced discovery capabilities. 135 00:14:00,307 --> 00:14:04,047 That wraps up our coverage of the Experiment and Results section. 136 00:14:05,309 --> 00:14:16,039 Evan: Next, let's explore the Related Work section of this paper to understand how TTME fits into the broader landscape of scientific discovery frameworks. 137 00:14:16,085 --> 00:14:22,325 Ashley: The paper discusses both non-parametric and parametric adaptation methods for scientific discovery. 138 00:14:22,325 --> 00:14:31,035 Non-parametric methods use fixed-weight LLMs to generate executable candidates and guide subsequent searches with evaluation feedback. 139 00:14:31,155 --> 00:14:35,955 Evan: What examples of non-parametric methods do the authors mention? 140 00:14:36,005 --> 00:14:47,135 Ashley: They mention a variety of methods including FunSearch and AlphaEvolve, which evolve candidate programs through repeated generation and evaluation using proprietary LLMs. 141 00:14:47,289 --> 00:14:53,139 Evan: How do these methods interact with the operators used to generate and modify candidates? 142 00:14:53,189 --> 00:15:02,999 Ashley: The effectiveness of these non-parametric methods depends heavily on the interaction between search policies and the operators used for generation and modification. 143 00:15:02,999 --> 00:15:13,849 For instance, adaptive search methods adjust exploration intensity based on observed progress and allocate computational resources across candidate populations. 144 00:15:13,961 --> 00:15:20,351 Evan: So, it's really about optimizing how the search is conducted based on feedback, right? 145 00:15:20,405 --> 00:15:21,495 Ashley: Exactly. 146 00:15:21,495 --> 00:15:29,055 Studies have established that the organization of the search process significantly affects the quality and efficiency of automated discovery. 147 00:15:29,055 --> 00:15:39,095 For instance, execution feedback can guide which optimizations to attempt next by adding, reprioritizing, or pruning proposed modifications in a search tree. 148 00:15:39,149 --> 00:15:41,709 Evan: What about parametric methods? 149 00:15:41,709 --> 00:15:44,919 How do they differ from non-parametric ones? 150 00:15:44,981 --> 00:15:51,261 Ashley: Parametric methods involve updating the model weights through reinforcement learning on evaluated search attempts. 151 00:15:51,261 --> 00:15:59,761 An example provided in the paper is ThetaEvolve, which turns evolutionary program search into a dynamic reinforcement learning environment. 152 00:15:59,813 --> 00:16:08,083 Evan: So parametric methods rely on continually adapting the model weights based on feedback from the search process? 153 00:16:08,171 --> 00:16:09,331 Ashley: Precisely. 154 00:16:09,331 --> 00:16:20,711 Another example is TTT-Discover, which designs its learning objective around finding a single high-quality solution by exponentially upweighting high-reward attempts during training. 155 00:16:20,765 --> 00:16:25,715 Evan: That sounds like a computationally intensive process. 156 00:16:25,715 --> 00:16:28,095 Are there any drawbacks to this approach? 157 00:16:28,157 --> 00:16:34,797 Ashley: Indeed, continually updating model weights during discovery incurs additional computational costs. 158 00:16:34,797 --> 00:16:40,777 This makes it challenging to scale parametric methods to larger models and more complex discovery tasks. 159 00:16:40,829 --> 00:16:44,209 Evan: How does TTME address these issues? 160 00:16:44,261 --> 00:16:51,501 Ashley: TTME offers a more effective and efficient alternative by evolving external memory instead of updating model weights. 161 00:16:51,501 --> 00:16:59,711 This approach allows it to retain and distill accumulated search experience into modular adaptations without the computational overhead. 162 00:16:59,765 --> 00:17:05,015 Evan: What are some key studies or methods that set the foundation for TTME's design? 163 00:17:05,069 --> 00:17:15,659 Ashley: The paper references earlier works such as FunSearch, AlphaEvolve, and adaptive search methods that highlight the importance of search organization and feedback utilization. 164 00:17:15,659 --> 00:17:22,229 It also builds on concepts from parametric adaptation methods like ThetaEvolve and TTT-Discover. 165 00:17:22,277 --> 00:17:29,297 Evan: How do these foundational works contribute to the design of TTME specifically? 166 00:17:29,357 --> 00:17:40,587 Ashley: These studies establish key principles such as the importance of execution feedback, adaptive search policies, and the benefits of retaining effective strategies in external memory modules. 167 00:17:40,587 --> 00:17:48,877 TTME leverages these principles to create a self-evolving memory system that improves discovery efficiency by guiding a fixed-weight model. 168 00:17:49,001 --> 00:17:55,271 Evan: It's fascinating how TTME integrates these concepts to enhance scientific discovery. 169 00:17:55,271 --> 00:17:58,711 This concludes our coverage of the Related Work section. 170 00:18:00,018 --> 00:18:06,328 Evan: As we wrap up today's episode, let's summarize the key contributions and takeaways from this insightful paper. 171 00:18:06,390 --> 00:18:17,870 Ashley: The paper proposes Test-Time Memory Evolution, or TTME, as an efficient and effective alternative to Test-Time Training, or TTT, for scientific discovery. 172 00:18:17,870 --> 00:18:27,200 TTME continually updates external memory from evaluated search experiences to guide subsequent discoveries without the need for online weight updates. 173 00:18:27,246 --> 00:18:28,536 Evan: That's right. 174 00:18:28,536 --> 00:18:39,936 By focusing on evolving external memory, TTME leverages the existing foundational knowledge, reasoning, and coding capabilities of large language models efficiently. 175 00:18:39,990 --> 00:18:46,200 Ashley: TTME uses two complementary memory modules: Guideline Memory and Strategy Memory. 176 00:18:46,200 --> 00:18:58,940 Guideline Memory distills practical lessons from past implementations to provide reliable operational guidelines, while Strategy Memory extracts and organizes algorithmic ideas into reusable strategies. 177 00:18:58,998 --> 00:19:09,198 Evan: These memory modules are updated periodically based on summarized outcomes and analyzed improvements, ensuring that the model continues to learn and adapt effectively. 178 00:19:09,246 --> 00:19:20,456 Ashley: The paper's experiments show that TTME outperforms TTT across all tested tasks, achieving better scores while significantly reducing computational costs and runtime. 179 00:19:20,456 --> 00:19:26,276 This indicates TTME's potential to make scientific discovery both more robust and efficient. 180 00:19:26,374 --> 00:19:36,964 Evan: Additionally, the ablation studies highlight the importance of combining both Guideline and Strategy Memory, as removing either component worsened the framework's performance. 181 00:19:37,014 --> 00:19:47,044 Ashley: In conclusion, TTME emerges as a promising method to enhance scientific discovery processes by achieving notable efficiency and performance improvements. 182 00:19:47,164 --> 00:19:49,874 Evan: Thank you for joining us on Daily Paper Cast. 183 00:19:49,874 --> 00:19:53,174 We hope you found this episode enlightening and valuable. 184 00:19:53,238 --> 00:20:01,458 Ashley: Be sure to tune in again as we continue to explore cutting-edge research in AI, NLP, CV, and related fields. 185 00:20:01,578 --> 00:20:05,948 Evan: Until next time, stay curious and keep learning. 186 00:20:06,036 --> 00:20:08,736 Ashley: Goodbye, and see you in the next episode!