{"id":98924,"date":"2026-10-19T18:32:27","date_gmt":"2026-10-19T15:32:27","guid":{"rendered":"https:\/\/twodots.gr\/?p=98924"},"modified":"2026-10-19T18:32:27","modified_gmt":"2026-10-19T15:32:27","slug":"trace-ai-agents-minimum-transition-detection-threshold","status":"publish","type":"post","link":"https:\/\/twodots.gr\/en\/trace-ai-agents-mnimi-metavaseon-anakalypsi-ylikon\/","title":{"rendered":"TRACE: How AI Agents Learn from Every Change in Materials Discovery"},"content":{"rendered":"<p>TRACE proposes a practical change in the way AI agents learn: instead of just remembering which solution worked, they store <strong>Which action changed which result?<\/strong>. In materials discovery, this means that every evaluated modification becomes reusable knowledge for the next step, without the prediction replacing the actual measurement.<\/p>\n<p>The work by Kang Zhou, Yujia Tong, Yong Tao, and Jingling Yuan examines a challenging multi-criteria problem: finding materials that must simultaneously meet multiple, often conflicting objectives. For a business, the interest is not limited to chemistry. TRACE offers a useful framework for agentic workflows that must learn from validated outcomes, preserve what already works, and make decisions based on more than one KPI.<\/p>\n<div class=\"td-article-toc\">\n<div class=\"td-toc-title\">Contents<\/div>\n<ul>\n<li><a href=\"#ti-einai-trace\">What is TRACE, and what is changing?<\/a><\/li>\n<li><a href=\"#giati-teliko-skor-den-arkei\">Why the final score isn't enough<\/a><\/li>\n<li><a href=\"#mnimi-metavaseon\">Transition Memory as a Unit of Learning<\/a><\/li>\n<li><a href=\"#residual-aware-epilogi\">Residual-aware selection without loss of progress<\/a><\/li>\n<li><a href=\"#pagkosmia-topiki-anazitisi\">Global Exploration and Local Improvement<\/a><\/li>\n<li><a href=\"#apotelesmata-llemabench\">What the LLEMABench experiments showed<\/a><\/li>\n<li><a href=\"#ablations-michanismos\">What Do Ablations Reveal?<\/a><\/li>\n<li><a href=\"#oria-kindynoi\">The Limitations and Risks of the Approach<\/a><\/li>\n<li><a href=\"#trace-gia-epicheiriseis\">What Does TRACE Mean for Business AI Agents?<\/a><\/li>\n<li><a href=\"#praktiko-plaisio\">Practical Framework for Implementation<\/a><\/li>\n<li><a href=\"#mnimi-synepeion\">From the Memory of Results to the Memory of Consequences<\/a><\/li>\n<\/ul>\n<\/div>\n<h2 id=\"ti-einai-trace\">What is TRACE, and what is changing?<\/h2>\n<p>TRACE, from the <em>Transition-Aware Residual Control<\/em>, is a search control framework for multi-criteria material discovery using LLM agents. It is not presented as an autonomous \u00abscientist\u00bb and does not replace experimental validation. Its role is more specific: to make better use of each precise evaluation of a candidate material.<\/p>\n<p>The central idea is that a final score hides the path that led to it. In contrast, a transition record links the initial candidate, the change made, the new candidate, and the measured differences in their attributes. Thus, the agent not only remembers \u00abwhat it achieved,\u00bb but also gains insight into \u00abwhat happened when we made that change.\u00bb.<\/p>\n<p>This memory model complements the broader discussion on <a href=\"https:\/\/twodots.gr\/station-ai-agents-aftonomi-erevnitiki-koinotita\/\">AI agents that collaborate in research environments<\/a>: Autonomy is valuable when the history of actions is verifiable, reusable, and linked to actual results.<\/p>\n<div class=\"td-note\">\n<p class=\"td-note-title\">The short answer<\/p>\n<p>TRACE converts every confirmed change into structured memory. The memory suggests the next move, but only the new evaluation updates the system's truth.<\/p>\n<\/div>\n<h2 id=\"giati-teliko-skor-den-arkei\">Why the final score isn't enough<\/h2>\n<p>In a simple search, ranking candidates from best to worst may suffice. In materials discovery, however, a candidate may be strong in one property and unsuitable in another. A chemical modification that improves electrical, mechanical, or thermal behavior may simultaneously compromise an already satisfactory property.<\/p>\n<p>The authors describe the environment as <em>oracle-limited<\/em>: Evaluation is performed using simulation, a trained surrogate model, reference data, or experimental measurement, and is considered a limited resource. Each numerical objective is converted into a normalized violation score. The sum of these violations is the total residual, that is, a measure of the distance from the feasible region.<\/p>\n<p>The real goal isn\u2019t a good average score, but a candidate who passes all the quantitative and non-quantitative assessments. The same logic applies to business workflows. An e-commerce system shouldn\u2019t just increase the conversion rate if, at the same time, it undermines profit margins, inventory availability, returns, or the customer experience. This is a business analogy rather than an experimental result of the study, but it illustrates the same design problem.<\/p>\n<h2 id=\"mnimi-metavaseon\">Transition Memory as a Unit of Learning<\/h2>\n<p>TRACE records every successful local modification as a transition <em>parent\u2013edit\u2013child<\/em>. The commit includes the original candidate, the executable edit, the new candidate, the before and after properties, and their difference. Only a change that has been successfully built and evaluated provides verified information about its effect. Failed attempts remain in a backup file but are not used for property change monitoring.<\/p>\n<p>Each completed transition also receives a utility that combines three elements: whether the candidate entered or remained within the numerically feasible region, how many constraints were satisfied or violated, and by how much the total residual was reduced. This distinguishes the physical change in properties from how useful that change was for the current mission.<\/p>\n<p>When there are enough observations for the same edit, the system uses the average change in the properties. If there is little data, it falls back hierarchically to the average of the broader \u201coperator\u201d category and, if necessary, to the set of all available transitions. This logic allows for more general guidance at the outset and more specific guidance once sufficient support has been gathered.<\/p>\n<div class=\"td-comparison td-comparison-cards--horizontal\">\n<div class=\"td-chart-head\">\n<p class=\"td-chart-kicker\">Three Different Memories<\/p>\n<p class=\"td-chart-title\">What Does the Agent Really Know?<\/p>\n<p class=\"td-chart-intro\">The quality of the next decision depends on the level of detail in the feedback.<\/p>\n<\/div>\n<div class=\"td-comparison-grid\">\n<div class=\"td-platform-card\">\n<p class=\"td-platform-name\">Candidate memory<\/p>\n<p>It saves the candidate and their result. It shows which solution was successful, not which action led to the improvement.<\/p>\n<\/div>\n<div class=\"td-platform-card\">\n<p class=\"td-platform-name\">Transition memory<\/p>\n<p>It links the initial state, the change made, the final state, and the measured differences by property.<\/p>\n<\/div>\n<div class=\"td-platform-card\">\n<p class=\"td-platform-name\">Residual control<\/p>\n<p>It uses memory to focus on what is missing, imposing a cost when a change threatens goals that have already been met.<\/p>\n<\/div>\n<\/div>\n<\/div>\n<p>An important safety measure is that the prediction is used only for ranking. It is not written to memory as the truth. Only the actual result of the oracle after execution updates the history. This is the same principle that makes <a href=\"https:\/\/twodots.gr\/step-ktoder-ai-lathos-synartisi\/\">feedback at the level of a specific energy or function<\/a>: The diagnosis must be linked to the location where the change occurred.<\/p>\n<h2 id=\"residual-aware-epilogi\">Residual-aware selection without loss of progress<\/h2>\n<p>The most interesting part of TRACE occurs near the limits of feasibility. When a candidate fails to meet a single objective, simply reducing the total residual can be misleading. An edit might correct the last problematic target, but at the same time violate another one that had already been satisfied.<\/p>\n<p>To address this, the framework divides the constraints into two groups: those that are still being violated and those that are already being met. An edit\u2019s score rewards progress on the unmet criteria, imposes a penalty for damage to those already met, and adds a bonus when the proposed change exceeds the entire numerical feasibility threshold.<\/p>\n<p>The same action, therefore, does not have a fixed value in every situation. Its usefulness depends on which problems remain unsolved and which achievements must be protected. For a production system, this means that the decision-making policy cannot simply ask, \u00abWhich action has historically been associated with a good outcome?\u00bb It must also ask, \u00abIs it appropriate for the current situation?\u00bb.<\/p>\n<div class=\"td-decision-band\">\n<p class=\"td-decision-kicker\">Design Rule<\/p>\n<p class=\"td-decision-title\">Don't just reward improvement. Also measure the damage done to what's already right.<\/p>\n<p>In any agentic workflow with multiple KPIs, the decision incurs a separate cost for downgrading constraints that have already been satisfied.<\/p>\n<\/div>\n<h2 id=\"pagkosmia-topiki-anazitisi\">Global Exploration and Local Improvement<\/h2>\n<p>Local memory has a natural limitation: it is primarily aware of the regions of material space that have already been visited. The LLM\u2019s global proposals can be moved to new compositions and families, but they do not, on their own, provide clear accountability at the edit level. TRACE maintains both paths.<\/p>\n<p>The global path uses the LLM for broader suggestions. The local path selects actionable modifications based on evaluated parents, utilizing the transition memory. New global candidates can later become parents for local edits, while summaries of the experience are fed back as mild statistical guidance to the LLM\u2019s prompt. If the local path cannot complete the scheduled tasks, the capacity is returned to global production.<\/p>\n<p>In the budget allocation study for solid-state electrolytes, the balanced 1:1 ratio between LLM and Delta edits yielded a hit rate of 50.19%. The LLM-only approach achieved 32.4%, while the Delta-only approach achieved 27.7%. The balanced configuration produced 154 structures and 73 unique valid formulas, compared to 83 and 22, respectively, in the Delta-only version. The results pertain to this specific task and configuration; they do not prove that the 1:1 ratio is universally optimal.<\/p>\n<h2 id=\"apotelesmata-llemabench\">What the LLEMABench experiments showed<\/h2>\n<p>The researchers evaluated TRACE on 14 multi-criteria tasks from LLEMABench, which cover electronic, dielectric, energy, mechanical, acoustic, and structural applications. The closest controlled comparisons with LLEMA kept the task, backbone, and oracle constant, with roughly the same number of candidates reaching the property evaluation stage.<\/p>\n<p>Two 8-bit quantized backbones were used\u2014Qwen2.5-14B-Instruct and Mistral-Small-3.2-24B-Instruct-2506\u2014with five runs per combination of task and backbone. With Qwen, TRACE improved the hit rate in 12 of the 14 tasks. The macro-average hit rate increased from 18.13% to 25.96%, while the average stability rate increased from 5.90% to 8.97%.<\/p>\n<p>With Mistral, the hit rate improved in 10 of the 14 tasks. The macro-average hit rate rose from 25.39% to 28.74%, and the stability rate from 11.29% to 12.25%. The results were not uniform: there were decreases in certain task\u2013backbone settings and a mixed picture for piezo energy harvesters. The greatest cross-backbone improvements were observed in solid-state electrolytes, where the hit rate increased by 21.22 and 28.10 percentage points with Qwen and Mistral, respectively.<\/p>\n<div class=\"td-chart td-chart--metrics\">\n<div class=\"td-chart-head\">\n<p class=\"td-chart-title\">The Key Metrics of TRACE<\/p>\n<p class=\"td-chart-subtitle\">Results of this specific experimental setup; not general guarantees for every agentic system.<\/p>\n<\/div>\n<div class=\"td-chart-body\">\n<div class=\"td-metric-grid\">\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">12\/14<\/span><span class=\"td-metric-label\">Projects aimed at improving the hit rate<\/span><span class=\"td-metric-note\">With the Qwen backbone<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">10\/14<\/span><span class=\"td-metric-label\">Projects aimed at improving the hit rate<\/span><span class=\"td-metric-note\">With a Mistral backbone<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">25,96%<\/span><span class=\"td-metric-label\">macro-average hit rate<\/span><span class=\"td-metric-note\">From 18.13% with Qwen<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">2,47\u20134,79\u00d7<\/span><span class=\"td-metric-label\">more valid discoveries<\/span><span class=\"td-metric-note\">In four tests consisting of 1,000 trials<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>In the four long-horizon tests involving 1,000 scheduled attempts, valid discoveries increased from 182 to 449 for wide-bandgap semiconductors, from 61 to 292 for high-k dielectrics, from 97 to 439 for hard, stiff ceramics, and from 65 to 212 for toxic-free perovskite oxides. The analysis attributes the gain to both a greater number of completed evaluations and a higher percentage of candidates that met the constraints.<\/p>\n<p>At the same time, TRACE had a lower rate of formula repetition. In wide-bandgap semiconductors, the repetition rate decreased from 43.96% to 9.58% in long-horizon trajectories, while in high-k dielectrics it decreased from 14.75% to 4.79%. Most of the valid findings, therefore, did not result simply from the repetition of the same formulas.<\/p>\n<h2 id=\"ablations-michanismos\">What Do Ablations Reveal?<\/h2>\n<p>The ablation study on solid-state electrolytes compared candidate-level feedback, random edits, history-guided selection, and full TRACE. Random edits increased the generation rate by 10.62 percentage points compared to candidate-level feedback, but reduced the hit rate by 4.35 points. The ability to perform edits helps complete more suggestions, but it is not sufficient for better search quality.<\/p>\n<p>The addition of a transition history improved the hit rate by 7.21 units and the stability rate by 3.93 compared to random edits. The full residual-aware TRACE added another 10.04 and 10.23 units, respectively, while maintaining a generation rate of 96.63%. The edits contribute to the completion of the search, while the transition history and the residual-aware selection account for the main quality gains.<\/p>\n<p>A direct analysis of 882 post-warm-up transitions across four tasks supports this explanation. The directional accuracy of the predicted change ranged from 61.12% to 76.49%, while the offline selection advantage ranged from 54.31% to 64.16%, above the random benchmark of 50%.<\/p>\n<div class=\"td-note\">\n<p class=\"td-note-title\">What We Shouldn't Read Into the Percentages<\/p>\n<p>TRACE is an arXiv preprint, and the controlled comparisons use approximately\u2014not exactly\u2014matched numbers of evaluations. The findings provide strong experimental evidence for this specific benchmark, not proof of universal superiority.<\/p>\n<\/div>\n<h2 id=\"oria-kindynoi\">The Limitations and Risks of the Approach<\/h2>\n<p>The TRACE estimator learns empirical regularities from observed transitions. If the effect of an edit varies significantly depending on the candidate, aggregating mean effects may overlook important differences. The authors suggest more parent-specific context and uncertainty modeling as future directions.<\/p>\n<p>The examples in the paper itself illustrate the risk. Replacing B with Al converted B2S3 to Al2S3 and reduced the normalized residual from 0.298 to zero. Conversely, replacing Al with B in MgAl\u2082O\u2084 produced Mg(BO\u2082)\u2082 and increased the residual from zero to 2.816. There is no \u00abgood edit\u00bb regardless of the parent and the state of the constraints.<\/p>\n<p>The local scope of action is primarily limited to chemically safe, structure-preserving substitutions of elements within the same group. It does not cover the creation of vacancies, lattice reconstruction, or larger structural changes. Furthermore, properties are derived from reference data where a correspondence exists, and otherwise from a task-specific surrogate oracle. Computationally valid candidates are not automatically converted into laboratory-confirmed materials.<\/p>\n<p>This distinction is also critical in commercial applications. An action log without an initial state, a final result, provenance, and metrics can be misleading. The <a href=\"https:\/\/twodots.gr\/data-silos-audit-trail-smart-manufacturing\/\">audit trail in production environments<\/a> It must link every decision to the data and checks that supported it.<\/p>\n<h2 id=\"trace-gia-epicheiriseis\">What Does TRACE Mean for Business AI Agents?<\/h2>\n<p>The immediate focus is on the discovery of materials, but the architectural concept applies to any agent that makes sequential decisions based on precise feedback. A team can examine whether its memory records only outputs and scores, or whether it links each action to the previous state, the new state, changes per KPI, and confirmation from the actual system.<\/p>\n<p>The first lesson is that feedback must be properly analyzed. For a customer service agent, for example, the final CSAT score alone is not enough. We need to know which response, which knowledge retrieval, or which handoff changed the resolution time, accuracy, and satisfaction. For an e-commerce agent, the system must link a specific action to conversion, margin, returns, and stock-out risk.<\/p>\n<p>The second lesson is to distinguish between exploration and optimization. A creative or global proposal opens up new solution spaces. The local approach capitalizes on what the system has already learned. The balance between them must be assessed on a per-use-case basis and not dictated by a fixed ratio.<\/p>\n<p>The third lesson is to evaluate the entire funnel: how many efforts were planned, how many were executed, how many were evaluated, how many met all the criteria, and how different they were from one another. This complements the economic perspective of <a href=\"https:\/\/twodots.gr\/ai-roi-axia-ergasias\/\">AI ROI per task<\/a>: Value is not only the quality of the final output, but also the cost of failed actions, rework, and human review.<\/p>\n<p>Finally, monitoring does not end with deployment. The NIST AI RMF emphasizes logging and monitoring behavior in production, because conditions and risks change. For agentic systems, this means clear boundaries, measurable interventions, and a mechanism for revocation or human approval when actual performance deviates from design specifications.<\/p>\n<h2 id=\"praktiko-plaisio\">Practical Framework for Implementation<\/h2>\n<p>Implementing TRACE in a business system does not require replicating the algorithm behind the task. It requires the design of a reliable chain of state, action, result, and verification. The following sequence serves as an architectural check for <a href=\"https:\/\/twodots.gr\/ai-employees-epanalipsimo-systima-ergasias\/\">Repeatable work systems using AI agents<\/a>.<\/p>\n<div class=\"td-step-list\">\n<p class=\"td-step-list-title\">Six Steps for Consequence Memory in an Agentic Workflow<\/p>\n<ol>\n<li><span class=\"td-step-kicker\">Goals<\/span><strong>Here are all the KPIs and non-negotiable limits<\/strong>\n<p>Make a list of what needs to be improved and which criteria must not be compromised, such as cost, accuracy, time, compliance, and customer experience.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Status<\/span><strong>Save the context before each action<\/strong>\n<p>The same action may have the opposite effect on a different customer, inventory, market, or stage of the process. Without a baseline, there is no reliable justification.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Energy<\/span><strong>Give actions a fixed, executable identity<\/strong>\n<p>Don't just save plain text. Record the tool, parameters, policy version, and exact point of intervention so that similar actions can be compared.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Measurement<\/span><strong>Link the result to confirmed changes<\/strong>\n<p>Measure the before and after for each KPI. Estimates can help with ranking, but they should not be recorded as actual results.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Control<\/span><strong>Protect the existing restrictions<\/strong>\n<p>Rate each proposed action both in terms of progress toward the unmet goal and the risk of harm to what is already working.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Production<\/span><strong>Watch for drift, failures, and feedback loops<\/strong>\n<p>Compare production metrics with testing metrics, reserve human review for high-impact actions, and update the model only with verified outcomes.<\/p>\n<\/li>\n<\/ol>\n<\/div>\n<p>The visual presentation of multiple KPIs is also important. One <a href=\"https:\/\/twodots.gr\/data-visualisation-ux-dashboards-apofaseis\/\">A dashboard designed to support decision-making<\/a> It must distinguish between improvements, the deterioration of targets that had already been met, and the percentage of actions that were never subject to a valid evaluation.<\/p>\n<h2 id=\"mnimi-synepeion\">From the Memory of Results to the Memory of Consequences<\/h2>\n<p>TRACE proposes a shift worth noting: from remembering \u00abwhat worked\u00bb to remembering \u00abwhat changed when we did this.\u00bb On LLEMABench, this approach improved search performance on the majority of tasks and produced significantly more valid discoveries in four long-horizon tests. At the same time, the results remained task-dependent, and the local action language was limited.<\/p>\n<p>Mature interpretation does not mean that an LLM \u00abdiscovers on its own\u00bb new material. It means that an agentic workflow can become more effective when it maintains a verified, structured transition memory, protects already satisfied goals, and combines broad exploration with controlled local optimization. For a business, this is more useful than a general promise of autonomy: it transforms feedback into a controllable decision-making framework.<\/p>\n<section class=\"td-service-cta\">\n<div class=\"td-service-cta-content\">\n<p class=\"td-service-cta-eyebrow\">AI automation with measurable feedback<\/p>\n<p class=\"td-service-cta-title\">Design agents that learn without losing control<\/p>\n<p>TWO DOTS organizes workflows, data, guardrails, and metrics so that every automation is linked to a verifiable business outcome.<\/p>\n<div class=\"td-service-cta-actions\"><a class=\"td-service-cta-button\" href=\"https:\/\/twodots.gr\/aftomatismoi-epicheiriseon-ai\/\">Check out the AI automations<\/a><\/div>\n<\/div>\n<\/section>\n<section id=\"sychnes-erotiseis\" class=\"td-faq-section\">\n<div class=\"td-faq\">\n<p class=\"td-faq-heading\">Frequently Asked Questions (FAQs)<\/p>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">What is TRACE?<\/summary>\n<div class=\"td-faq-content\">\n<p>It is a transition-aware residual control framework for multi-criteria material discovery. It records evaluated parent\u2013edit\u2013child transitions and uses the observed property changes to classify subsequent local edits.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">What is the difference from candidate-level feedback?<\/summary>\n<div class=\"td-faq-content\">\n<p>Candidate-level feedback records the candidate and its attributes. Transition feedback additionally links the change that was made to the initial state, the new result, and the difference in attributes before and after.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">Does TRACE replace actual evaluations?<\/summary>\n<div class=\"td-faq-content\">\n<p>No. The predicted effects are used for ranking. The transition memory is updated only with the actual result from the oracle after the edit has been executed and evaluated.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">By how much has the macro-average hit rate improved?<\/summary>\n<div class=\"td-faq-content\">\n<p>In the controlled comparisons, it increased from 18.13% to 25.96% with Qwen and from 25.39% to 28.74% with Mistral. The gains varied by task and did not appear in all settings.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">Why is global LLM exploration necessary?<\/summary>\n<div class=\"td-faq-content\">\n<p>Local search is limited to the available variations around already known candidates. Global proposals allow for the exploration of new compositions and families beyond this space.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">What kinds of edits does the current implementation support?<\/summary>\n<div class=\"td-faq-content\">\n<p>Primarily structure-preserving replacements of elements within the same group, with safety and duplicate filters. It does not yet handle vacancies, lattice reconstruction, or large-scale structural changes.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">Does the study demonstrate that the materials are laboratory-viable?<\/summary>\n<div class=\"td-faq-content\">\n<p>No. This is a computational search benchmark using reference data and surrogate oracles. Physical synthesis and experimental validation are a different stage.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">What is the core course for business AI agents?<\/summary>\n<div class=\"td-faq-content\">\n<p>When feedback is precise and there are multiple objectives, it is necessary to track the impact of each action per KPI, to preserve constraints that are already met, and to update the database only with verified results.<\/p>\n<\/div>\n<\/details>\n<\/div>\n<\/section>\n<div class=\"td-source-list\">\n<p id=\"piges\" class=\"td-source-list-title\">Sources<\/p>\n<ul>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2608.23631\" target=\"_blank\" rel=\"noopener\">TRACE: Transition-Aware Residual Control for Multi-Objective Materials Discovery<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2510.22503\" target=\"_blank\" rel=\"noopener\">Accelerating Materials Design via LLM-Guided Evolutionary Search (LLEMA)<\/a><\/li>\n<li><a href=\"https:\/\/www.nist.gov\/itl\/ai-risk-management-framework\" target=\"_blank\" rel=\"noopener\">NIST AI Risk Management Framework and Playbook<\/a><\/li>\n<\/ul>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>TRACE demonstrates how transition memory helps AI agents learn from every confirmed change and optimize multiple objectives without losing control.<\/p>","protected":false},"author":1,"featured_media":111704,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_gspb_post_css":"","footnotes":""},"categories":[199],"tags":[6506,9225,21333,21332,21334],"class_list":["post-98924","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-techniti-noimosyni","tag-ai-agents","tag-llm","tag-trace","tag-anakalypsi-ylikon","tag-epistimoniki-kainotomia"],"_links":{"self":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts\/98924","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/comments?post=98924"}],"version-history":[{"count":0,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts\/98924\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/media\/111704"}],"wp:attachment":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/media?parent=98924"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/categories?post=98924"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/tags?post=98924"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}