Prabhav Singh (@psingh522) header
Prabhav Singh (@psingh522) profile picture

PhD @UTAustin | Interested in computational linguistics, specifically uncertainty/ambiguity in discourse.

Austin, TX
Joined October 2024
Photo by Prabhav Singh (@psingh522)Photo by Prabhav Singh (@psingh522)
Super excited to share that I will be starting my PhD in Computer Science at @UTAustin/@UTCompSci this fall! Thrilled to be working with @EliasEskin and @jessyjli on topics like uncertainty quantification, attribution, and clarification interactions in LMs. Needless to say, I have a lot of people to thank at @jhuclsp (continued in 🧵)
15
10
2
101
9,138
Prabhav Singh retweeted
🚨Agentic Multimodal ICL: Can an MLLM learn to build its own in-context examples: what to look for, what to keep, and when it has enough? We post-train MLLM agents to do this with tailored tools. ActICL beats similarity retrieval by +11.6 points, and boosts frontier models including GPT 5.6! Key takeaways: 1️⃣ We cast ICL retrieval/context construction as an agentic task: the agent searches a demonstration bank with retrieval goals it writes itself, reads promising candidates, curates its context, and commits it, all under token budgets. 2️⃣ We train ActICL agents with rejection-sampling SFT followed by joint RL over acquisition and answering. Across 7 benchmarks and 3 backbones, ActICL beats the strongest trained baseline by +4.4 points while using only ~3.5 demonstrations on average. 3️⃣ ActICL contexts transfer: a 9B ActICL agent builds contexts that raise GPT-5.6-Sol by +3.5 points over standard retrieval, and it beats a Gemini3.5-Flash-powered agentic workflow on all 9 model–benchmark pairs we test. 👇🧵
3
9
3
34
3,458
Prabhav Singh retweeted
Deployed agents pile up execution traces. The question is whether we can automatically refine skills based on existing rollouts. Introducing SkillRefiner: Offline skill refinement from the traces you already have! 🧵
1
16
3
66
5,539
Prabhav Singh retweeted
🚨 We introduce the problem of persistent calibration for trustworthy, continually learning models and find substantial headroom even for confidence estimators that appear well-calibrated at a single point in time. As a model gains more knowledge, we would like its confidence to track that change in knowledge, without requiring repeated re-calibration. To test how existing methods do, we evaluate persistent calibration on checkpoints of open models, asking whether confidence estimators trained on earlier checkpoints can generalize to later ones, whose knowledge has changed. To measure the dependence of confidence on knowledge, we evaluate calibration on knowledge contrast sets: subsets containing questions that one checkpoint answers correctly and another checkpoint answers incorrectly, reflecting a change in knowledge. Key findings: 1️⃣ Both inference-time and fine-tuning methods fall short on contrast-set calibration, even when they are well-calibrated on the full set. We quantify the headroom with oracle methods trained on future checkpoints – we call this an oracle because persistent calibration requires learning to estimate confidence in a way that generalizes to future checkpoints, whereas the oracle trains directly on them. 2️⃣ Multi-checkpoint training helps in both the non-oracle and oracle settings, suggesting an avenue for identifying confidence features that remain robust across changing knowledge. 3️⃣ Our analysis shows that two confidence adapters trained on different checkpoints can perform similarly on one checkpoint, but extremely differently on another checkpoint. This suggests that there is a vast space of possible confidence functions that are well-calibrated on a given checkpoint, out of which only some rely on meta-knowledge features that would generalize to other checkpoints, presenting a challenge for persistent calibration.
1
19
2
30
3,413
I tried to not post this but the rage bait got to me. The whole point of discovering good music is that it's fun and rewarding. I don't think I would get the same happiness from "discovering" new music from AI that just personalises its generated stuff to match everything else I like. Not to mention that this harms basically everyone involved: the listener and (more so) the person making music (for who this probably takes out all of the creative joy of making new music).
I am not joking when I say this. Spotify, Apple Music, and streaming in general is over. AI music is now very, very good. You just don't know yet because all AI music circulating today is from ~2-3 generations ago. Suno v6 pro et al has solved infinite personalized music.
1
144
Prabhav Singh retweeted
When do LLM agents develop new languages that we can’t understand? Lots of recent news about this, based mostly on anecdata from a single run. We study language emergence more rigorously, finding key factors like LLM strength, access to scratchpad messages, and pressure for efficiency. Studying the languages themselves, we find they are morphologically productive, compositional, and can be transmitted to new agents, including agents backed by weaker models, even ones not able to develop language on their own. To study language emergence systematically, we developed a new platform, GlossoGen, which lets us design controlled, sandboxed multi-agent scenarios with different initial conditions and dynamics. We instantiate one such scenario and use it to study open and closed-weight models across many runs. Key takeaways: 1️⃣ Sufficiently strong models, under pressure to communicate efficiently and with access to a postmortem scratchpad, develop new languages. 2️⃣ Languages are compositional and morphologically productive. 3️⃣ Languages can be transmitted to new learners who observe them being used without seeing their construction. 4️⃣ Even models that are not strong enough to construct languages can learn to use them. Agents take an active role in learning languages, with new agents repairing failed conversations via targeted queries. More details in our paper below, including implications for safety/monitorability, cumulative cultural evolution, and linguistics. 🧵👇
3
38
7
137
12,131
Replying to @ben_vandurme
@ben_vandurme and I are recruiting multiple postdoc fellows at JHU. We're looking for candidates w/ strong record in language models, reasoning, coding agents, and/or AI for science. Interested candidates should send their CV and a brief summary of their research interests to [email protected] / [email protected]. Please reshare for visibility. 🙏
31
3
52
13,606
Prabhav Singh retweeted
NLAs are claimed to verbalize model activations. But can they faithfully interpret steered activations? In our latest paper, we show that steering moves activations into non-invertible regions; and almost surely, no prompt maps to steered activations! NLAs fail to interpret steered activation states faithfully, supporting our results! ↓ @anqi_liu33 @DanielKhashabi xitter.gq/AnthropicAI/status/205…
New Anthropic research: Natural Language Autoencoders. Models like Claude talk in words but think in numbers. The numbers—called activations—encode Claude’s thoughts, but not in a language we can read. Here, we train Claude to translate its activations into human-readable text.
19
98
16
607
89,661
Prabhav Singh retweeted
New 📄: we replicate XTR, a multi-vector retrieval method that makes ColBERT faster by avoiding its expensive step of gathering full document embeddings XTR is not a free lunch over ColBERT, but its training objective is useful for modern efficient engines like PLAID and WARP 👇🏼
2
27
5
64
12,509
Prabhav Singh retweeted
“Everyone can code now!” Dude, no one can code now.
193
1,225
79
19,905
339,973
Really excited about this project from our team! AgentOdyssey allows you to create open-world, long-horizon environments on demand— which can then be used to evaluate LLM agents on exploration, memory, and planning over time.
Introducing AgentOdyssey — an open-ended, long-horizon text game generation engine for 𝐭𝐞𝐬𝐭-𝐭𝐢𝐦𝐞 𝐜𝐨𝐧𝐭𝐢𝐧𝐮𝐚𝐥 𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐚𝐠𝐞𝐧𝐭𝐬. Real-world agents cannot have a boundary between training and testing: they must learn continuously from interaction with the world at test time. AgentOdyssey is designed to study five key abilities that make this possible: 🧭 𝐞𝐱𝐩𝐥𝐨𝐫𝐚𝐭𝐢𝐨𝐧, 🧠 𝐞𝐩𝐢𝐬𝐨𝐝𝐢𝐜 𝐦𝐞𝐦𝐨𝐫𝐲, 🌍 𝐰𝐨𝐫𝐥𝐝 𝐤𝐧𝐨𝐰𝐥𝐞𝐝𝐠𝐞 𝐚𝐜𝐪𝐮𝐢𝐬𝐢𝐭𝐢𝐨𝐧, 🛠️ 𝐬𝐤𝐢𝐥𝐥 𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠, and 🎯 𝐥𝐨𝐧𝐠-𝐡𝐨𝐫𝐢𝐳𝐨𝐧 𝐩𝐥𝐚𝐧𝐧𝐢𝐧𝐠. Just `pip install agentodyssey` and start generating games and evaluating your agents and LLMs! 🌐 Project website: agentodyssey.github.io/
5
34
2,888
Prabhav Singh retweeted
✈️ Will be in Rio for ICLR #iclr2026 for: 📄 calibration w/ self-generated distractors 📄 symbolic world models in stochastic envs 📄 code understanding via runtime exec traces 📄 scene graphs for art description 📄 benchmark for trustworthy LLM/VLMs 📢 keynote + panel at the Workshop on Multi-Agent Learning & Gen AI Come say hi! 🧵👇
1
7
40
2,769
It has been a wonderful journey at JHU and I couldn't have dreamt of a better place to do my master's. CLSP has some of the friendliest people who are always open to collaboration and some amazing courses taught by amazing professors! A special thank you to @adveisner from whom I've learned more about the field (and other random topics) than I could have ever asked for! Huge thank you also to @DanielKhashabi, @jvillalba1980, and @JessSorrell who I have had the honor of working with and learning from during my 2 years here.
1
6
833
Finally, a big thank you also to all the amazing PhD students at @jhuclsp who helped me a lot with valuable advice during the application season! I had to rewrite my SoP many times, but each time was worth it ❤️ Hoping to keep collaborating with the wonderful people here, and excited to meet the incredible folks in Austin!
2
506
Excited to share our new paper: "Are Finer Citations Always Better?" 🔍 In attributed generation, models output responses & citations in a single pass. Humans prefer sentence-level citations for fast verification, but if models can't generate them faithfully, quality degrades.
1
10
15
1,360
Prabhav Singh retweeted
The annual MASC conference (non-archival, student led, local conference) will be at Howard this year. Please consider submitting work that you’ve sent elsewhere in the past year. It’s a great networking opportunity!
Exciting opportunity! Mid-Atlantic Student Colloquium on AI, Language & Learning (MASC-ALL 2026) is happening April 25 at @HowardU Submit by April 13! mascsll.org Present your work + connect with others in NLP, speech & AI #MASCALL #NLP #AI #Research
3
7
1,004
Exciting opportunity! Mid-Atlantic Student Colloquium on AI, Language & Learning (MASC-ALL 2026) is happening April 25 at @HowardU Submit by April 13! mascsll.org Present your work + connect with others in NLP, speech & AI #MASCALL #NLP #AI #Research
13
1
24
2,969
Prabhav Singh retweeted
New blog post: the state of AI safety in four fake graphs.
63
144
37
1,389
667,547