Andreas Göldi
Last Name
Göldi
First name
Andreas
Email
andreas.goeldi@unisg.ch
8 Ergebnisse
Gerade angezeigt 1 - 8 von 8
- third-party-metrics-blockedthird-party-metrics-cookies.consent-settings
Item type:Veröffentlichung, Efficient Management of LLM-Based Coaching Agents' Reasoning While Maintaining Interaction Quality and Speed(2025); ; Ungar, LyleLLM-based agents improve upon standalone LLMs, which are optimized for immediate intent-satisfaction, by allowing the pursuit of more extended objectives, such as helping users over the long term. To do so, LLM-based agents need to reason before responding. For complex tasks like personalized coaching, this reasoning can be informed by adding relevant information at key moments, shifting it in the desired direction. However, the pursuit of objectives beyond interaction quality may compromise this very quality. Moreover, as the depth and informativeness of reasoning increase, so do the number of tokens required, leading to higher latency and cost. This study investigates how an LLM-based coaching agent can adjust its reasoning depth using a discrepancy mechanism that signals how much reasoning effort to allocate based on how well the objective is being met. Our discrepancy-based mechanism constrains reasoning to better align with alternative objectives, reducing cost roughly tenfold while minimally impacting interaction quality.cris-layout.advanced-attachment.dc.type:conference paperJournal:Proceedings of the CHI Conference on Human Factors in Computing Systems (CHI ’25)metric-badges.scopusCitation 6 - third-party-metrics-blockedthird-party-metrics-cookies.consent-settings
Item type:Veröffentlichung, Flexibility in Chatbot Identity Perception(2024)In recent years, chatbots have become much more prevalent and capable. Users think of chatbots as machines but treat them as social actors. The perception of chatbots as machines and as social actors each brings with it both benefits and drawbacks. If it were possible to switch flexibly between the two perspectives, users could retain benefits while avoiding drawbacks. For example, they could commit to the chatbot as a social actor to enjoy its flattery, and still dismiss its utterances as mere machine productions when they trigger a feeling of doubt. We varied reminders about the chatbot’s identity to observe which would help most to keep enjoying its exaggerated flattery. We reminded the groups that the chatbot was a machine, a social actor, or that the user was free to switch their perspective on the chatbot’s identity. Only if users already believed they could switch was the last reminder impactful.cris-layout.advanced-attachment.dc.type:conference paperJournal:Americas Conference on Information Systems (AMCIS) - third-party-metrics-blockedthird-party-metrics-cookies.consent-settings
Item type:Veröffentlichung, Chatbot Agents Displaying Non-factive Reasoning Enhance Expectation Confirmation(2024); Current guidelines suggest setting low expectations for chatbots for continued engagement. That is of little use if only high expectations entice users. Such are raised by recent innovations, chatbot agents which generate final responses not directly via Large Language Models but after multiple intermediate generations of text as reasoning. This differentiates chatbot agents from other chatbots. Since these intermediate steps are generated, what impact has displaying them? This study aims to assess the impact of displaying these intermediate reasoning steps, conceptualizing them as enabling mindreading in the semi-literal sense, namely reading the internal reasoning generated by the agent. In 3 studies (N=280), we examine whether non-factive displays, which present reasoning as belief rather than knowledge, enhance user expectation confirmation. Results show that non-factive reasoning displays significantly improved confirmation. Therefore, displaying non-factive reasoning not only differentiates these agents in competitive markets but also improves user interaction.cris-layout.advanced-attachment.dc.type:conference paperJournal:International Conference on Information Systems (ICIS) - third-party-metrics-blockedthird-party-metrics-cookies.consent-settings
Item type:Veröffentlichung, Insert-expansions for Large Language Model Agents(2024); AI agents use language generation for internal reasoning and interacting with users. Agents often employ tools beyond language generation, such as calculators or search, to further augment these capabilities. We focus on how such tools can give the agent too much external context, diverting it from the user’s original intent. According to Conversation Analysis, human-human dialogue often uses ”insert-expansion” - inserted utterances for clarification - to resolve ambiguities. Building on this, we introduce a ”user-as-a-tool” approach, enabling the AI agent to solicit clarification from the user while still reasoning, thereby realigning it with the user’s intent. Initial evidence shows that our approach has benefits for conversational recommendation systems. We present a novel interaction method and empirical findings that enhance the user’s role in guiding agent reasoning. This research is especially relevant as AI agents become increasingly common, and holds significance for optimizing the human-chatbot interaction loop.cris-layout.advanced-attachment.dc.type:conference paperJournal:Americas Conference on Information Systems (AMICS) - third-party-metrics-blockedthird-party-metrics-cookies.consent-settings
Item type:Veröffentlichung, Making Sense of Large Language Model-Based AI Agents(2024); Large Language Models (LLMs) have had major impact in society even though most LLM applications use single model calls to generate output. Recent innovations have uncovered that multiple chained calls tend to produce better results. Even more impactful is the discovery that these chains do not need to be predefined. LLM-based AI agents use frameworks to generate written intermediate reasoning that decides which steps to take next and when to return with a final output. LLM-based AI agents can use external tools like search engines, calculators, code engines, etc. to gather information and act on the world. Developments in this area are rapid and potentially consequential. However, it is difficult to keep apace with the developments. To address this, we introduce a typology grounded in recent research that provides a structured framework for understanding LLM-based agents, facilitating proactive engagement with future developments.cris-layout.advanced-attachment.dc.type:conference paperJournal:International Conference on Information Systems (ICIS) - third-party-metrics-blockedthird-party-metrics-cookies.consent-settings
Item type:Veröffentlichung, Intelligent Support Engages Writers Through Relevant Cognitive Processes(2024-05-11); ; metric-badges.scopusCitation 24 - third-party-metrics-blockedthird-party-metrics-cookies.consent-settings
Item type:Veröffentlichung, Insert-expansions for Tool-enabled Conversational Agents(2023-07-04); This paper delves into an advanced implementation of Chain-of-Thought-Prompting in Large Lan- guage Models, focusing on the use of tools (or "plug-ins") within the explicit reasoning paths generated by this prompting method. We find that tool-enabled conversational agents often become sidetracked, as additional context from tools like search engines or calculators diverts from original user intents. To address this, we explore a concept wherein the user becomes the tool, providing necessary details and refining their requests. Through Conversation Analysis, we characterize this interaction as insert-expansion — an intermediary conversation designed to facilitate the preferred response. We explore possibilities arising from this ’user-as-a-tool’ approach in two empirical studies using direct comparison, and find benefits in the recommendation domain.cris-layout.advanced-attachment.dc.type:conference paper - third-party-metrics-blockedthird-party-metrics-cookies.consent-settings
Item type:Veröffentlichung, WHERETO FOR AUTOMATED COACHING CONVERSATION: STRUCTURED INTERVENTION OR ADAPTIVE GENERATION?(2023); In an age of lifelong learning, it is important that adult learners can effectively use their motivation and resources to reach their learning goals. In conversation, coaches can intervene to promote learning goal attainment by using behavioural change techniques (BCTs). In a coaching chatbot, such techniques can be ordered in an established, structured way to good effect. With recent technological advances, chatbot responses can be generated adaptively; this means that BCTs can be applied in an adaptive but less structured way. It is yet unclear whether this latter form of configuring coaching interventions is viable, how they compare to more established structured interventions, and whether both methods can be combined. For the purpose of answering this, we propose a 2x2 experimental design with the two intervention types as factors and goal attainment as the dependent variable. Results will indicate avenues for automating skilled conversation including choice of technology.cris-layout.advanced-attachment.dc.type:conference paperJournal:European Conference on Information Systems (ECIS)