Thomas J. Fan

Most engineers know this: stochastic systems don’t fail loudly. They fail quietly, in small inconsistencies that add up. We saw it when testing LLMs for alert investigations. Same input → different paths → different outcomes. Sometimes a CTI query was skipped. Sometimes the severity rating changed. Sometimes the investigation was shorter, with missing context. Not “wrong,” but incomplete. And in a SOC, incomplete is wrong. Our study (18,000 investigations across 180 real alerts) showed just how much this variability matters: → Canonical paths rarely exceeded 75% consistency → Complex alerts produced almost as many unique paths as attempts → Even identical alerts sometimes led to conflicting conclusions This is the fundamental limit of stochastic intelligence. At Qevlar, we address it by separating reasoning from orchestration. LLMs bring the analytical depth. Our graph engine enforces the structure that guarantees every critical step is executed. Because resilience in security comes from certainty. We break it down in detail in our latest research: https://lnkd.in/eUNysiaw

48

6 Comments

Our colleagues at Lawrence Livermore National Laboratory and ETH Zürich have introduced SPADA, "Spatial Dataflow Architecture Programming Language." This programming language provides precise control over data placement, dataflow patterns, and asynchronous operations while abstracting architecture-specific design. As part of this work, the team implemented a compiler targeting CSL and the Cerebras SDK, and demonstrated that SPADA can serve as both a high-level programming interface and an intermediate representation for DSLs, enabling developers to express complex parallel patterns for the Cerebras architecture. Tal Ben-Nun #IAmCerebras https://lnkd.in/gkUnru3G

60

1 Comment

I just published Vol. 140 of "Top Information Retrieval Papers of the Week" on Substack. My Substack newsletter features the 7-10 most notable research papers on information retrieval (including recommender systems, search & ranking, etc.) from each week, with a brief summary, and links to the paper/codebase. This week’s newsletter highlights the following research work: 📚 Deep GraphRAG: A Balanced Approach to Hierarchical Retrieval and Adaptive Integration, from Ant Group 📚 Progressive Reinforcement Learning with Semantic IDs for Negative Feedback Modeling, from Alibaba 📚 Tree-Structured Reasoning for Robust Multi-Hop Question Answering with RAG, from Shi et al. 📚 Sparse Autoencoders for Interpretable and Steerable Collaborative Filtering, from Spišák et al. 📚 Unifying Long-Sequence Modeling and Feature Interaction for Industrial-Scale CTR Prediction, from ByteDance 📚 Confidence-Guided Pruning for Efficient Multi-Hop Retrieval-Augmented Reasoning, from Jiao et al. 📚 Learning to Retrieve for Agentic Search, from Liu et al. 📚 Predicting Context Utility and Answer Quality in RAG, from the University of Glasgow 📚 Interpreting and Steering Popularity Bias Through Sparse Autoencoder Neurons in Recommender Systems, from TU Delft 📚 Aligning Document Ranking with Generator Preferences for RAG, from Fan et al. #InformationRetrieval #ResearchPapers #CuratedContent #Newsletter  #substack

33

2 Comments

RL often throws away useful signal at intermediate steps, or as Karpathy put it, it's like "sucking supervision through a straw." MiniMax M2.5 solves this with per-token process rewards. The result is frontier coding performance at least 1/10th the cost of closed source. Alex Ker breaks down how this mechanism works and how M2.5 excels in general knowledge work. Read about it here: https://lnkd.in/eqZkN_5p

30

Google researchers just published something quietly interesting: a method for teaching LLMs to reason like a Bayesian. Here's the problem they were solving. Standard LLMs are bad at updating their beliefs mid-conversation. You give them new information, and instead of revising their model of the world, they often just... incorporate it awkwardly. They don't maintain a running probability distribution. They don't apply Bayes' rule. They pattern-match to what a helpful response looks like. The Google team tested this with a flight recommendation task. A Bayesian assistant — one that explicitly tracks and updates user preferences over multiple turns — dramatically outperformed a standard LLM at making accurate recommendations as the conversation evolved. Their fix: Bayesian teaching. Train the LLM to imitate the behavior of an optimal Bayesian system during simulated interactions. Not to understand probability theory, but to behave as if it does. This is a subtle but important distinction. You're not giving the model a math lesson. You're shaping its behavior through imitation of a better reasoner. The broader implication: a lot of LLM failures in multi-step tasks come down to this exact problem — the model doesn't know what it doesn't know, and it doesn't update well when it's wrong. Bayesian teaching might be one of the more practical paths to fixing that. What reasoning failure mode do you run into most often in multi-turn LLM interactions? https://lnkd.in/gxMQjEQC

9

1 Comment

Exciting news 🎉 : two papers accepted at EMNLP W-NUT 2026 1. Topic Matching in the Wild: Benchmark and Lessons from Real-World ASR Transcripts Accepted, top 50%. This one was an amazing collaboration with some genuinely brilliant people and my collaborators at Dialpad: Xiliang Zhu, Irvin Cardoza, and David Rossouw. We looked at what actually happens when you try to match conversation topics on raw production speech recognition output, not the clean text most benchmarks assume. (This was an exciting work and unlocked a lot of new findings that are ongoing for Dialpad) 2. The Curse of Multilinguality in Lexical Normalization (Strong accept) Saman (Sam) Rahbar (individual author) Done entirely as independent research. It asks a question that sounds simple and turns out not to be: when you train one multilingual model, is more languages always better? #EMNLP2026 #NLP #MachineLearning #SpeechRecognition #MultilingualNLP #Research

94

7 Comments

We found that fine-tuning a language model doesn't just add new features — it reorganizes existing ones. Inside a Marcus Aurelius LoRA, the real adaptation lives in co-activation clusters of shared features, not individual interpretable units. Clusters produce causal effects 10× beyond any single feature. Link in comments.

1

1 Comment

The paper introduces Tree Training, a novel approach to train agentic large language models (LLMs) whose rollout trajectories form branching trees rather than straight lines. Current pipelines treat each branch as an independent linear sequence, causing redundant computation for shared prefixes. The authors address this inefficiency with two key techniques: Tree Packing, which reuses computation of shared prefix segments across branches; and Gradient Restoration, which correctly propagates gradients through reused segments. Empirical results show up to a 3.9× speed‑up in training time on open‑source models, significantly improving efficiency in agentic LLM supervised (SFT) or reinforcement‑learning training. https://lnkd.in/gbRcMUs6

1