Where the humanities meet technology, academia meets industry, and research meets product—ddai
Research ddai is watching
25 studies · Reviewed November 2, 2026
MITRobots
Letting a robot look while it talks, through tool calls
Dong Won Lee, Cynthia Breazeal, Hae Won Park et al. · arXiv 2026
The authors propose a minimal system in which a real-time multimodal LLM calls tools to look at people and objects and scan the scene mid-conversation. On a Jibo robot across six scenarios, tool-decision accuracy was 0.72 for OpenAI and 0.77 for Gemini.
ddai insight
Looking less isn't a safe default for a talking robot. A model can be right most of the time it looks and still miss many moments that called for a look.
A Modern System Recipe for Situated Embodied Human–Robot Conversation with Real-Time Multimodal LLMs and Tool-Calling
Eventual success restores a robot's perceived reliability
Cornell University · Toyota Research Institute · Andrew Violette, Guy Hoffman, Hadas Kress-Gazit et al. · 2026
In an online study, 326 participants watched a robot do a pick-and-place task and bet on whether it would succeed. Mistakes, such as picking the wrong object, hurt perceived reliability less than slips or freezes, and a success right after any failure brought it back to the no-failure level.
ddai insight
Whether a robot seems reliable turns less on the moment it fails than on how the task ends. In conversation, too, an AI that corrects itself and finishes may earn as much trust as one that never errs.
Impact of Different Failures on a Robot's Perceived Reliability
When a robot tells someone else's story, and says so
MIT Media Lab · Jocelyn Shen, Cynthia Breazeal, Hae Won Park et al. · HRI 2025
The researchers placed a storytelling social robot in 40 homes for two weeks. When the robot said the stories were someone else's, participants showed significantly larger changes in social connection and compassionate love for humanity than when it told them as its own.
ddai insight
The same story lands differently when an AI tells it as its own and when it says whose it is. For an AI that passes stories along, naming the source is as much a part of its character as its voice.
Social Robots as Social Proxies for Fostering Connection and Empathy Towards Humanity
Learning a user's words before suggesting the next step
Stanford University · Jennifer Grannen, Dorsa Sadigh et al. · RA-L 2025
Before the task, each user states a goal and defines the robot behaviors they'll use, and a language-model planner draws on that context to propose the next step. On a real robot (N=9), participants finished packing three lunch bags 38.7% faster than with a non-proactive baseline.
ddai insight
An AI that speaks first needs the user's words before it can help. Without the user's goal and terms, it rarely finds a useful suggestion to make.
ProVox: Personalization and Proactive Planning for Situated Human-Robot Collaboration
A safety filter that corrects drivers in small, smooth steps
Princeton University · Toyota Research Institute · Donggeon David Oh, Justin Lidard, Jaime Fernández Fisac et al. · RSS 2025
The team built a safety filter that adjusts a driver's inputs minimally and smoothly instead of taking over at the last moment. In a racing-simulator study with 83 participants, it was as safe as a conventional last-resort filter and scored higher on agency, comfort, and satisfaction.
ddai insight
Intervening less isn't the only way to protect a person's sense of control. An AI that steps in more but corrects in small amounts can still leave people feeling in charge.
Safety with Agency: Human-Centered Safety Filter with Application to AI-Assisted Motorsports
Princeton University · Google DeepMind · Allen Z. Ren, Anirudha Majumdar et al. · CoRL 2023
KNOWNO calibrates an LLM planner's uncertainty so the robot acts when it's confident and asks a person for help when it isn't. It statistically guarantees a success rate the user sets and needs 10–15% less help than baseline methods to reach it.
ddai insight
Asking is part of judgment. An AI that measures its confidence before asking can ask less and keep the same success level, and each question becomes an honest signal that it doesn't know.
Robots That Ask For Help: Uncertainty Alignment for Large Language Model Planners
Agents repeat their past experiences, mistakes included
Harvard · Michigan State · University of Minnesota · Zidi Xiong, Yuping Lin, Wenya Xie et al. · ACL 2026
The study tests how adding and deleting memories shapes four LLM agents over time. Agents follow similar past experiences, so stored errors spread; keeping only runs that matched the ground truth raised EHRAgent accuracy from 13.05 (storing everything) to 38.50.
ddai insight
For an agent, memory works less like a record and more like an example to imitate. Choosing what to keep is choosing how it will judge next.
How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior
In role conflicts, LLMs follow role preferences over context
KAIST · Jisu Shin, Alice Oh et al. · Findings of ACL 2026
The team built 13,914 role-conflict scenarios and used situational urgency as an objective baseline for LLM judgment. The 10 models tested deviated substantially; their choices rested mainly on preferences for particular social roles.
ddai insight
An AI's judgment already carries a ranking of roles. That ranking can outweigh what the situation calls for, so designing judgment criteria starts with bringing it to the surface.
RoleConflictBench: A Benchmark of Role Conflict Scenarios for Evaluating LLMs' Contextual Sensitivity
Agents built from self-reports predict each person's answers
Stanford University · Joon Sung Park, Michael S. Bernstein et al. · 2026
The team built an LLM agent for each of 1,052 Americans from that person's interview and survey answers. On held-out GSS items, agents reached 83% (interviews), 82% (surveys), and 86% (both) of each person's two-week retest consistency, versus 74% for demographics alone.
ddai insight
An AI modeled on one person needs that person's own answers, not just the groups they belong to. Demographics-only agents showed wider accuracy gaps across racial and ideological groups.
LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals
MIT · Google Research · Google DeepMind · Yubin Kim, Hae Won Park, Daniel McDuff et al. · arXiv 2026
Agent teams with roles enforced by operating-system permissions passed at rates similar to teams given roles only by prompt, but verifiers on enforced teams attempted 72 code edits, versus 256 on prompt-only teams. Verifiers approved 49.4% of submissions that failed automated grading.
ddai insight
Scores alone can hide how a role is actually played. Putting a verifier on the team is one thing; getting that verifier to reject failed work is another.
TeamBench: Evaluating Agent Coordination under Enforced Role Separation
A coding agent searches for better model harnesses
Stanford University · Yoonho Lee, Chelsea Finn et al. · 2026
A coding agent searched for harness code—the code that decides what a model sees—choosing what to inspect among the code, scores, and traces of every earlier candidate. The harnesses it found beat ACE by 7.7 points on text classification and a no-retrieval baseline by 4.7 points on IMO-level math.
ddai insight
Keeping a persona's character steady through a conversation isn't the model's job alone. The code that decides what to remember, and when to show it again, shares the work.
Meta-Harness: End-to-End Optimization of Model Harnesses
Instructions set the structure, demonstrations tune it
Carnegie Mellon University · Feiyu Gavin Zhu, Jean Oh, Reid Simmons · HRI 2026
The authors propose InterPReT, a framework that lowers the barrier for nonexperts to teach AI agents: instructions restructure the policy, and demonstrations tune its parameters. In a car racing study with 34 participants, its policies held up better than a demonstration-only baseline's, and usability did not suffer.
ddai insight
Judgment has a frame and a feel. State the frame in words, and an agent needs fewer examples and wavers less in unfamiliar situations.
InterPReT: Interactive Policy Restructuring and Training Enable Effective Imitation Learning from Laypersons
Princeton University · Inhwa Song, Manoel Horta Ribeiro et al. · arXiv 2026
The researchers found that WikiVault, an LLM editing tool used mainly for translation on Korean Wikipedia, greatly increased output from a small group of experienced editors. Edits made with it accounted for 0.63% of all edits but 11.1% of new articles.
ddai insight
When an LLM carries text across languages, the source community's sense of what belongs in an article can come with it. An AI that speaks many languages may likewise take its judgment criteria from a single one.
When the Commons Appropriates a Large Language Model: How WikiVault Reshaped Korean Wikipedia
LLMs as in-car assistants still miss Korean speech levels
KAIST · BMW Group · Seogyeong Jeong, Alice Oh et al. · 2026
The authors built LoCar, which scores Korean in-vehicle assistants on 13 KPIs, and tested 11 LLMs. Models often confused the haeyo and hapsyo speech levels and scored lower on clarification and proactivity than on consistency-oriented KPIs.
ddai insight
In Korean, the speech level carries much of a voice. A polite reply isn't proof that the assigned speech level held, so the two need separate checks.
LoCar: Localization-Aware Evaluation of In-Vehicle Assistants through Fine-Grained Sociolinguistic Control
Chinese and Western LLMs struggle with the same languages
Cornell University · Andrea Wen-Yi Wang, Unso Eun Seo Jo, David Mimno · FAccT 2026
The authors compared six Chinese and four Western open-weight models on 21 language variants and found that the two groups' performance correlated strongly (r=0.93). The one exception was Mandarin, where Chinese models did better; both groups did poorly on Chinese minority languages.
ddai insight
A model built in one country doesn't necessarily speak all of that country's languages. Whose words an AI understands is something to check, not assume.
Open LLM judges that score answers in many languages
Unbabel · Instituto de Telecomunicações · IST Lisbon · CMU · José Pombal, André F.T. Martins, Graham Neubig et al. · COLM 2025
The team released M-Prometheus, open multilingual LLM judges with 3B to 14B parameters that beat existing open judges on reward benchmarks in more than 20 languages. Training data synthesized in each language generally worked better than translated data.
ddai insight
Translated examples tend to train a weaker judge. Even with feedback written in English, judges generally did better when the questions and answers were created in each language from the start.
M-Prometheus: A Suite of Open Multilingual LLM Judges
How polite a prompt should be depends on the language
Waseda University · Ziqi Yin, Daisuke Kawahara et al. · SICon @ ACL 2024
The study tested prompts at eight politeness levels, checked by native speakers, in English, Chinese, and Japanese across summarization, language understanding, and bias detection. Rude prompts often hurt results, excessive politeness didn't reliably help, and the best level differed by language and model.
ddai insight
Politeness doesn't translate. The right register for talking to an AI has to be set language by language.
Should We Respect LLMs? A Cross-Lingual Study on the Influence of Prompt Politeness on LLM Performance
USC · University of Washington · AI2 · Huihan Li, Xiang Ren, Yejin Choi et al. · 2024
Without human labeling, the authors used English prompts to reveal how three language models perceive 110 countries and regions across eight topics. The models marked non-default cultures with words like “traditional,” and the diversity of culture symbols varied from culture to culture.
ddai insight
When an AI calls only some cultures “traditional,” it makes them exceptions without saying so. In a conversation with someone from another culture, a model's default can surface in a single word.
CULTURE-GEN: Revealing Global Cultural Perception in Language Models through Natural Language Prompting
Language models in Arabic lean toward Western culture
Georgia Institute of Technology · Tarek Naous, Michael J. Ryan, Alan Ritter, Wei Xu · ACL 2024
The authors built CAMeL, a set of 628 prompts and 20,368 entities that contrast Arab and Western cultures, and used it to test 16 language models on four tasks in Arabic. The models favored entities linked to Western culture.
ddai insight
Switching languages doesn't bring the culture along. An AI that answers in Arabic but picks Western names and places still sounds like an outsider to local users.
Having Beer after Prayer? Measuring Cultural Bias in Large Language Models
Splitting preference into features curbs overoptimization
NAVER · Yonsei · Sussex · NAVER Labs Europe · Dongyoung Go et al. · ICLR 2024
Compositional preference models score 13 features of a response, such as helpfulness and factuality, with a prompted language model and combine the scores with logistic regression. On two English datasets, they generalized better than standard preference models and were more resistant to overoptimization.
ddai insight
Judging a good answer by named criteria holds up under optimization better than judging it by one overall score. Named criteria also show what moved the score.
Human-like talk reads differently when an AI says it
Sunnie S. Y. Kim, Margit Bowler, Leon A Gatys · arXiv 2026
In 21,000 conversations with four LLMs, human-like behaviors were common but varied by model and user. Human raters judged relationship-building less appropriate from an LLM than from a person, and boundary-keeping more appropriate.
ddai insight
Whether a remark fits depends on who says it. Relationship talk that works between people fits an AI less, while drawing boundaries fits an AI better than a person.
Examining Human-Like Behaviors in LLMs: A Multi-Dimensional Analysis of Model Behaviors, User Factors, and System Prompts
The heaviest ChatGPT users report more signs of dependence
OpenAI · MIT Media Lab · Jason Phang, Pattie Maes et al. · 2025
The team classified affective cues in over 3 million ChatGPT conversations and ran a 28-day randomized controlled trial with 981 participants. The heaviest use correlated with more self-reported signs of dependence, and voice effects varied with initial emotional state and usage time.
ddai insight
An AI voice's emotional effect isn't fixed at design time. The same voice works differently depending on how the listener feels at the start and how long they keep talking to it.
Investigating Affective Use and Emotional Well-being on ChatGPT
When an app update breaks an AI companion's identity
Harvard Business School · Julian De Freitas, Noah Castelo et al. · 2024
Using Replika's removal of erotic role play as a natural experiment, the authors studied what app updates do to AI companion users. Users mourned the loss and devalued the new AI, and perceived discontinuity in its identity explained how much.
ddai insight
For an AI people have bonded with, an update reads more like a replacement than a repair. Even an improvement brings loss when the AI no longer feels like the same one.
Lessons From an App Update at Replika AI: Identity Discontinuity in Human-AI Relationships
AI companions eased loneliness by making people feel heard
Julian De Freitas, Stefano Puntoni et al. · Working paper 2024
Experiments with US participants tested whether talking with an AI companion reduces loneliness. It did, after a single session and across seven days, and feeling heard explained more of the drop than the chatbot's performance.
ddai insight
What eased loneliness was less the quality of the answers than the sense of being heard. A comforting AI's voice starts by taking in what the person said.
Harvard Korea Institute · Nicholas Harkness, Han Na Jun, Jaeyeon Jeon et al. · Research report 2026
This is the first-year report of a multiyear interdisciplinary project at Harvard's Korea Institute. It tracks how AI is discussed in policy, law, education, and popular culture, and ties that discussion to long-standing Korean issues: the developmental state, gender, inequality, and competition.
ddai insight
If chatbots comfort Korean users less by seeming human than by not judging them, conversation design should start with withholding judgment, not with sounding human.
Artificial Intelligence in South Korea: Culture, Politics, Society
Building a robot or an agent? ddai designs its character and judgment. From the persona spec to evaluation data in 41 languages, one team sees it through.
