If you serve as an EMNLP (i.e., ARR May) Area Chair, you may receive results from an automatic check for AI-generated fake citations, powered by GPTZero. It seems that automatically checking for fake citations is becoming standard practice at AI conferences. I am not sure whether this trend will have a positive or negative impact. For example, if a paper was only recently accepted to a conference and someone cites the conference name in the reference, could it be mistakenly flagged as AI-generated? I am somewhat worried about false positives like this.

This might be the first open-source memory architecture to cover associative recall. The rest stop at similarity search. (T-Mem, EMNLP 2026) Most memory systems retrieve by similarity. T-Mem predicts when a memory will matter, before you ask. → write: memories ship with a trigger at store time, basically "when will this come up again?" → read: at query time, triggers pull in memories by situation, even with zero keyword overlap. → architecture layer. plugs into any LLM without fine-tuning or model changes. On LoCoMo-Plus (strips keyword overlap, tests pure associative recall), mainstream systems drop 28 to 50 points. T-Mem drops 5.45. paper: code: (MIT)

🎉 Excited to share that our paper "VEX-Bench: Benchmarking LLM Agents for Assessing Exploitability of Software Supply Chain Vulnerabilities” has been accepted to EMNLP 2026! This work is a collaboration between @PurdueCS and multiple teams from @RedHat (AI Engineering, Ecosystem Engineering, and Product Security). Software supply chains are an increasingly important security challenge. While tools can identify vulnerable dependencies, determining whether a vulnerability is actually exploitable in a downstream application often requires substantial cross-repository reasoning and security expertise. In this work, we introduce VEX-Bench, a benchmark for evaluating LLM agents on this problem: - 75 real-world vulnerability cases mined from GitHub - Expert-labeled across Python, Java, and Go - Evaluation of 9 models across 3 agent harnesses - Going beyond binary “exploitable / not exploitable” decisions to evaluate the reasoning behind exploitability Our results show that frontier LLM agents can perform well on binary vulnerability assessment, but fine-grained exploitability reasoning remains significantly more challenging. This points to an important direction for AI-assisted software security: it’s not enough for an agent to identify that something is vulnerable. We need to understand why it is exploitable, whether it actually matters in context, and how confidently we can trust the assessment. 🙌 Huge thanks to all the co-authors and collaborators who made this possible! @tian_yi_zhang, Jiahao Shi, Yifeng Di, Ilan Pinto, Edward Tsien, Hongjiao Zhang, Ronit Dey, Ilona Shishov, Gal Netanel, Zvi Grinberg, Vladimir Belousov, Bat-Zion Rotman 📄 Paper: #EMNLP2026# #AI# #LLM# #LLMAgents# #Cybersecurity# #SoftwareSupplyChain# #OpenSource# #PurdueCS# #RedHat#

Can web agents learn through imagination? DynaWeb: Model-Based Reinforcement Learning of Web Agents has been accepted to EMNLP 2026 Main Conference. Training web agents with online RL means costly, risky interaction with the live web. DynaWeb learns a Web World Model instead, letting agents dream web interactions and train on imagined rollouts. - Learns naturalistic web transitions from agent actions - Generates imagined trajectories in a synthetic web environment - Interleaves imagined rollouts with real expert trajectories - Consistent gains for open-source agents on WebArena and WebVoyager Imagination as a scalable path toward online agentic RL. This is part of our ongoing work at Gradient on distributed and efficient RL: training capable agents without the cost, latency, and risk of centralizing them. Paper: Code:

Some friends told me that this year’s ARR May cycle, i.e., EMNLP 2026, seems to have much lower average scores than previous years. If your scores are not very high, don’t panic too much. I’ve seen papers with a 2.5 meta-review score in ARR still get accepted by top ACL-family conferences when the AC liked the work. Just move on, keep calm, and fingers crossed.

Fun fact: your agent is quietly wrong about the past. Outdated facts aren't deleted, just shadowed to the back. RoMem fixes that. Drop-in reranker, three lines to plug in. The key ideas: > Spin, don't delete Outdated facts rotate out of alignment at query time. Nothing erased. > Some relations expire faster A pretrained gate scores how fast each relation changes. born_in → near zero. works_at → fast. Zero-shot to unseen relations. > Time as rotation, not a label One continuous function. Any timestamp. Including dates never seen in training. 2–3x better at temporal reasoning, with zero drop in static memory. Joint research from Edinburgh, LIGHTSPEED and St Andrews. EMNLP 2026 Main. Paper: Code: