Joined June 2023
- Tweets61
- Following400
- Followers297
- Likes460
Yuqing Yang @ COLM 2026 retweeted
We find that gradients are useful for detecting implicit reward hacking that CoT monitoring can miss.
#COLM2026 poster (Thursday afternoon, Imperial 19)
LLMs can hide reward hacking behind convincing CoTs, making it difficult to detect these behaviors by reading CoT alone. This concern becomes even more pressing for latent or looped LMs.
Our method, Gradient Fingerprint (GRIFT), computes gradients of a model's CoT and compresses them into compact representations. These gradient fingerprints help separate hacking from non-hacking trajectories.
We also show using GRIFT to filter reasoning traces for rejection fine-tuning can suppress reward hacking.
Yuqing Yang @ COLM 2026 retweeted
I’ll be at the Bay Area for @COLM_conf next week! I’m on the job market this year and I’ll be presenting 3 posters on:
1️⃣ How different models learn similar number representations. convergent-evolution.github.…
2️⃣ Token-level off-policy learning. arxiv.org/abs/2607.17524
3️⃣ SAEs for Reasoning. arxiv.org/abs/2506.09967
Yuqing Yang @ COLM 2026 retweeted
Introducing VaSE: Value-Aware Stochastic KV Cache Eviction.
Reasoning models think in CoT, bloating the KV cache. Eviction caps memory but suffers capability drop. VaSE is a training-free recipe that cuts that cost: keep large-magnitude value states, evict stochastically.
Yuqing Yang @ COLM 2026 retweeted
Excited to share that I’ve started @GoogleResearch as a student researcher today. I'll be working on tabular foundation models. Come and chat if you are around at Google or at the Bay Area.
Yuqing Yang @ COLM 2026 retweeted
The future risk of computer-use agents won’t come only from malicious prompts. It will come from agents that can flawlessly follow normal instructions straight into harm.
Introducing 𝐎𝐒-𝐁𝐥𝐢𝐧𝐝: a realistic but overlooked setting where every task begins with a benign user instruction, yet the harmfulness only emerges as the agent acts in the environment.
Yuqing Yang @ COLM 2026 retweeted
New paper: Convergent Evolution: How Different Language Models Learn Similar Number Representations.
Language models, classical word embeddings, and even raw token frequencies all develop the same Fourier features for numbers. But only some develop the underlying structure. 🧵
Yuqing Yang @ COLM 2026 retweeted
After three papers on Fourier features in LLMs, I think there's a principle worth naming. How should we do science on an LLM?
It corresponds to the existential questions:
> who am I? ↔ the phenomenon.
> where do I come from? ↔ the emergence.
> where am I going? ↔ the use.
🧵
Yuqing Yang @ COLM 2026 retweeted
Frontier LLMs don't debug, they regenerate.
We built PDB to measure that gap, GPT-5.1-Codex pass unit tests >76% of the time, but touch only <45% of the right lines.
Even Claude Code touches only ~50%.
📄 Paper: arxiv.org/abs/2604.17338
🌐 Project: precise-debugging-benchmark.…
Yuqing Yang @ COLM 2026 retweeted
Announcing 🔭✨Hubble, a suite of open-source LLMs to advance the study of memorization!
Pretrained models up to 8B params, with controlled insertion of texts (e.g., book passages, biographies, test sets, and more!) designed to emulate key memorization risks 🧵
Yuqing Yang @ COLM 2026 retweeted
# 🚨 4B open-recipe model beats Claude-4-Opus
🔓 100% open data, recipe, model weights and code.
Introducing Polaris✨--a post-training recipe for scaling RL on advanced reasoning models.
🥳 Check out how we boost open-recipe reasoning models to incredible performance levels (65 → 79 on AIME25) through RL training on open-source data and academic-level resources.
📑Notion: honorable-payment-890.notion…
📗Blog post: hkunlp.github.io/blog/2025/P…
🤗Model & data: huggingface.co/POLARIS-Proje…
💻Code: github.com/ChenxinAn-fdu/POL…
Yuqing Yang @ COLM 2026 retweeted
There’s been hot debate about (The Illusion of) The Illusion of Thinking.
My take: it’s not that models can’t reason — they just aren’t perfect at long-form generation yet.
We eval reasoning models on LongProc benchmark (requiring generating 8K CoTs, see thread). Reasoning models actually outperform instruction-tuned ones, showing the benefit of long CoT training for long outputs. Still, there’s plenty of room for improvement.
See details in our post 👉 princeton-pli.github.io/Long…
🤔Now most LLMs have >= 128K context sizes, but are they good at generating long outputs, such as writing 8K token chain-of-thought for a planning problem？
🔔Introducing LongProc (Long Procedural Generation), a new benchmark with 6 diverse tasks that challenge LLMs to synthesize highly dispersed information and generate long, structured outputs.
Yuqing Yang @ COLM 2026 retweeted
🧵 Recent studies show LLMs can self-improve their responses when given external feedback. But how effectively can they incorporate it?
We tested this systematically—and found they can't fully integrate feedback, even when the feedback is high-quality and backed by ground-truth.
Yuqing Yang @ COLM 2026 retweeted
🚨 We discovered a surprising side effect of Reinforcement Finetuning (RFT): it makes LLMs more confidently wrong on unanswerable questions.
We call this the hallucination tax: a drop in refusal behavior that leads to overconfident hallucinations.
🧵 1/n
Yuqing Yang @ COLM 2026 retweeted
Textual steering vectors can improve visual understanding in multimodal LLMs!
You can extract steering vectors via any interpretability toolkit you like -- SAEs, MeanShift, Probes -- and apply them to image or text tokens (or both) of Multimodal LLMs.
And They Steer!
Yuqing Yang @ COLM 2026 retweeted
Want to know what your LLM don’t know?
This is how 👇
Preprint: arxiv.org/abs/2503.23361
Code: github.com/uscnlp-lime/SEA
Yuqing Yang @ COLM 2026 retweeted
Billion-parameter LLMs still struggle with simple arithmetic?
📞 FoNE (Fourier Number Embedding) tackles this problem.
By mapping numbers directly into Fourier space, it bypasses tokenization and significantly improves numerical accuracy with better efficiency and accuracy.
Yuqing Yang @ COLM 2026 retweeted
Running your model on multiple GPUs but often found the speed not satisfiable? We introduce Ladder-residual, a parallelism-aware architecture modification that makes 70B Llama with tensor parallelism ~30% faster!
Work done at @togethercompute. Co-1st author with @MayankMish98 and mentored by @ben_athi, @tri_dao
arxiv.org/pdf/2501.06589 🧵[1/7]
Yuqing Yang @ COLM 2026 retweeted
Come join the #NeurIPS2024 poster session and discuss whether language models can learn to skip steps in reasoning!
🗓Dec 12, Thursday, 11:00 am - 2:00 pm
📍East Exhibit Hall A-C #2900
Feel free to stop by and say hi! I am actively seeking Summer 2025 internship opportunities!
🤔Can LMs learn to skip steps to improve reasoning efficiency while maintaining accuracy?
✅The answer is Yes! In our #NeurIPS 2024 work, we show this behavior boosts efficiency, maintains accuracy, and even enhances generalization in OOD scenarios!
🚀arxiv.org/pdf/2411.01855
🧵⬇️
Yuqing Yang @ COLM 2026 retweeted
I'll present a poster for Lifelong ICL and Task Haystack at #NeurIPS2024!
⏰ Wednesday 11am-2pm
📍 East Exhibit Hall A-C #2802
📜 arxiv.org/abs/2407.16695
My co-first author @xiaoyue02_xu is applying to PhD programs and I am looking jobs in industry! Happy to connect at NeurIPS!
Introducing 𝗟𝗶𝗳𝗲𝗹𝗼𝗻𝗴 𝗜𝗖𝗟 and 𝗧𝗮𝘀𝗸 𝗛𝗮𝘆𝘀𝘁𝗮𝗰𝗸, a new approach for evaluating long-context LMs, featuring ever-changing task streams that controllably fill the context window, and NIAH-style visualization for easy diagnosis.
📜 arxiv.org/abs/2407.16695
🧵
Yuqing Yang @ COLM 2026 retweeted
Drop by our poster tomorrow!! @emnlpmeeting #EMNLP2024
Nov 12 (Tue) at 11:00-12:30
Session: 02
Sub-session: Generation
Looking forward to chatting with everyone!
How can LLMs 🤖 find all 𝙚𝙨𝙨𝙚𝙣𝙩𝙞𝙖𝙡 facts to write about a topic? 🤔
In our work, we leverage the planning capabilities of LLMs to guide retrieval of fine-grained facts for improved grounding of response.
See our #EMNLP2024 paper: aclanthology.org/2024.emnlp-…
[1/N]