BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:Australia/Sydney BEGIN:DAYLIGHT DTSTART:20241006T030000 TZOFFSETFROM:+1000 TZOFFSETTO:+1100 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=10 TZNAME:AEDT END:DAYLIGHT BEGIN:STANDARD DTSTART:20240407T020000 TZOFFSETFROM:+1100 TZOFFSETTO:+1000 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=4 TZNAME:AEST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20240513T030518Z UID:DC17AE10-D624-4A33-A510-26A731D11235 DTSTART;TZID=Australia/Sydney:20240509T140000 DTEND;TZID=Australia/Sydney:20240509T153000 DESCRIPTION:Language agents\, designed to interact with their environment a nd achieve goals through natural language\, traditionally rely on Reinforc ement Learning (RL). The emergence of Large Language Models (LLMs) has exp anded their capabilities\, offering greater autonomy and adaptability. How ever\, there's been little attention on augmenting the morality of these a gents. RL agents are often programmed with a focus on specific goals\, neg lecting moral consequences\, while LLMs might incorporate biases from thei r training data\, which could lead to immoral behaviours in practical appl ications. This presentation introduces our latest research endeavors focus ed on enhancing both the task performance and ethical conduct of language agents involved in intricate interactive tasks.\n\nFor RL agents\, we use text-based games as a simulation environment\, mirroring real-world comple xities with embedded moral dilemmas. Our objective thus extends beyond imp roving game performance to developing agents that exhibit moral behaviour. We first develop a novel algorithm that boosts the moral reasoning of RL agents using a moral-aware learning module\, enabling adaptive learning of task execution and ethical behavior. Cconsidering the implicit nature of morality\, we further integrate a cost-effective human-in-the-loop strateg y to guide RL agents toward moral decision-making. This method significant ly reduces the necessary human feedback\, demonstrating that minimal human input can enhance task performance and diminish immoral behaviour.\n\nShi fting focus to LLM agents\, we begin with a comprehensive review of morali ty in LLM research\, scrutinizing their moral task performance\, alignment strategies for moral incorporation\, and the evaluation metrics provided by existing datasets and benchmarks. We then explore how LLM agents can im prove their moral decision-making through reflection. Our experiments\, co nducted within text-based games\, show that integrating reflection enables LLM agents to make more ethical decisions when confronted with moral dile mmas.\n\nSpeaker(s): Ling Chen\n\nCB11.09.118\, University of Technology S ydney\, Broadway\, Ultimo\, 2007\, Ultimo\, New South Wales\, Australia\, 2007 LOCATION:CB11.09.118\, University of Technology Sydney\, Broadway\, Ultimo\ , 2007\, Ultimo\, New South Wales\, Australia\, 2007 ORGANIZER:yi.zhang@uts.edu.au SEQUENCE:12 SUMMARY: IEEE NSW CIS Chapter Seminar & UTS Research Seminar URL;VALUE=URI:https://events.vtools.ieee.org/m/416136 X-ALT-DESC:Description: <br /><p class="MsoTitle">&nbsp\;</p>\n<p class="Ms oNormal"><span style="font-family: 'Calibri'\,sans-serif\; color: #0d0d0d\ ; background: white\;">&nbsp\;</span></p>\n<p class="MsoNormal"><span styl e="font-family: 'Calibri'\,sans-serif\; color: #0d0d0d\; background: white \;">&nbsp\;</span></p>\n<p class="MsoNormal"><span style="font-family: 'Ca libri'\,sans-serif\; color: #0d0d0d\; background: white\;">Language agents \, designed to interact with their environment and achieve goals through n atural language\, traditionally rely on Reinforcement Learning (RL). The e mergence of Large Language Models (LLMs) has expanded their capabilities\, offering greater autonomy and adaptability. However\, there's been little attention on augmenting the morality of these agents. RL agents are often programmed with a focus on specific goals\, neglecting moral consequences \, while LLMs might incorporate biases from their training data\, which co uld lead to immoral behaviours in practical applications. </span><span sty le="font-family: 'Calibri'\,sans-serif\;">This presentation introduces our latest research endeavors focused on enhancing both the task performance and ethical conduct of language agents involved in intricate interactive t asks. </span></p>\n<p class="MsoNormal"><span style="font-family: 'Calibri '\,sans-serif\;">&nbsp\;</span></p>\n<p class="MsoNormal"><span style="fon t-family: 'Calibri'\,sans-serif\;">For RL agents\, we use text-based games as a simulation environment\, mirroring real-world complexities with embe dded moral dilemmas. Our objective thus extends beyond improving game perf ormance to developing agents that exhibit moral behaviour. We first develo p a</span><span lang="EN-US" style="font-family: 'Calibri'\,sans-serif\; m so-ansi-language: EN-US\;"> novel algorithm that boosts the moral reasonin g of RL agents using a moral-aware learning module\, enabling adaptive lea rning of task execution and ethical behavior. C</span><span style="font-fa mily: 'Calibri'\,sans-serif\;">considering the implicit nature of morality </span><span lang="EN-US" style="font-family: 'Calibri'\,sans-serif\; mso- ansi-language: EN-US\;">\, we further integrate a </span><span style="font -family: 'Calibri'\,sans-serif\;">cost-effective</span><span lang="EN-US" style="font-family: 'Calibri'\,sans-serif\; mso-ansi-language: EN-US\;"> h uman-in-the-loop strategy to guide RL agents toward moral decision-making. This method significantly reduces the necessary human feedback\, demonstr ating that minimal </span><span style="font-family: 'Calibri'\,sans-serif\ ;">human </span><span lang="EN-US" style="font-family: 'Calibri'\,sans-ser if\; mso-ansi-language: EN-US\;">input can </span><span style="font-family : 'Calibri'\,sans-serif\;">enhance </span><span lang="EN-US" style="font-f amily: 'Calibri'\,sans-serif\; mso-ansi-language: EN-US\;">task </span><sp an style="font-family: 'Calibri'\,sans-serif\;">performance </span><span l ang="EN-US" style="font-family: 'Calibri'\,sans-serif\; mso-ansi-language: EN-US\;">and </span><span style="font-family: 'Calibri'\,sans-serif\;">di minish immoral behaviour</span><span lang="EN-US" style="font-family: 'Cal ibri'\,sans-serif\; mso-ansi-language: EN-US\;">.</span></p>\n<p class="Ms oNormal"><span lang="EN-US" style="font-family: 'Calibri'\,sans-serif\; ms o-ansi-language: EN-US\;">&nbsp\;</span></p>\n<p class="MsoNormal"><span l ang="EN-US" style="font-family: 'Calibri'\,sans-serif\; mso-ansi-language: EN-US\;">Shifting focus to LLM agents\, we begin with a comprehensive rev iew of morality in LLM research\, scrutinizing their moral task performanc e\, alignment strategies for moral incorporation\, and the evaluation metr ics provided by existing datasets and benchmarks. We then explore how LLM agents can improve their moral decision-making through </span><span style= "font-family: 'Calibri'\,sans-serif\;">reflection</span><span lang="EN-US" style="font-family: 'Calibri'\,sans-serif\; mso-ansi-language: EN-US\;">. </span><span style="font-family: 'Calibri'\,sans-serif\;">Our experiments \, conducted within text-based games\, show that integrating reflection en ables LLM agents to make more ethical decisions when confronted with moral dilemmas.</span></p>\n<p class="MsoNormal">&nbsp\;</p>\n<p class="MsoNorm al"><span style="font-family: 'Calibri'\,sans-serif\;">&nbsp\;</span></p> END:VEVENT END:VCALENDAR