Tim Rocktäschel

Great couple of weeks for the lab, and for Andrew Bender, PhD in particular! Two of Andrew's papers diving deep into the meaning of neural alpha oscillations are now published. The first, out in PNAS, is "Differential representations of spatial location by aperiodic and alpha oscillatory activity in working memory" [1]. There's been a ton of super cool work in the last decade or so showing that the amplitude and distribution of visual cortical alpha rhythms encode the contents and spatial location of information held in working memory. Our lab has shown that our ability measure neural oscillations is confounded by non-oscillatory aperiodic activity, so we were curious how much of this encoding is *really* oscillatory. Analyzing seven different EEG datasets from Ed Awh and Ed Vogel's labs, we found that, after correcting for aperiodic activity, alpha encoding actually *improved*! And that time-resolved aperiodic activity seemed more to encode the stimulus while it was on the screen. We're now looking much more deeply into this idea that sensory information that often comes into our brains is aperiodic, and encoded that way, but then somehow transforms into an oscillatory neural code for maintenance. The second paper, in the Journal of Cognitive Neuroscience, is "Resting-state Alpha and Mu Rhythms Change Shape across Development But Lack Diagnostic Sensitivity for Attention-Deficit/Hyperactivity Disorder and Autism" [2]. LinkedIn's word limits keep me from being able to dive too deeply into this one but WOW did Andrew and Natalie Schaworonkow do an incredible amount of beautiful, technically sophisticated work. They analyzed EEG data from thousands of children, classified alpha band activity either as visual cortical alpha or sensorimotor mu based on their spatial topography, and then quantified the fine-scale waveform features of both. (By the way, the pre-prints of both these papers are on freely available on bioRxiv if you can't access the published versions.) [1] https://lnkd.in/eN9rxdZZ [2] https://lnkd.in/eNCjyQYh

97

2 Comments

NeurIPS 2026 call for papers is live now. Make your AI agents get to work. Papers won't write themselves. Tell them to hurry! /s

1

Automated software unit testing (including oracles) using a collaborative, multi-agent architecture based on Large Language Models "Hallucination to Consensus: Multi-Agent LLMs for End-to-End JUnit Test Generation", Qinghua Xu, Guancheng Wang, Lionel Briand, Kui Liu This paper, led by Qinghua Xu in my team at the Lero Centre, was accepted for publication in ACM Transactions on Software Engineering and Methodology. The project was supported by Huawei. All code and data are publicly available. We propose CANDOR, a novel end-to-end, prompt engineering-based LLM framework for automated unit test generation in Java. CANDOR orchestrates multiple specialized LLM agents to collaboratively generate complete JUnit tests, including both high-quality test prefixes and accurate oracles. To mitigate the notorious hallucinations in LLMs and improve oracle correctness, we introduce a novel strategy that engages multiple reasoning LLMs in a panel discussion and generates accurate oracles based on consensus. Preprint: https://lnkd.in/enu5adY3

67

3 Comments

Reconfigurable intelligent surfaces (RIS/IRS) are likely to be key enablers of enhanced wireless communications going forward. A key problem, of determining the "coding patterns" of the RIS elements was solved by us in a recent paper https://lnkd.in/d6HPfu4n. Today, we're happy to release the source codes for our algorithms for the community to build on: https://lnkd.in/d3GHihm8 [With Radha Krishna Ganti and Sai Sanjay Narayanan]

127

1 Comment

🚀 Really excited to see this go out. Over the past several months, our focus has been on building not just models, but the infrastructure layer required to make LLMs truly work in the Indian context—across training, alignment, and evaluation. With the open-sourcing of ForgeLM and IndicEvalHarness, we’re sharing two important pieces of that stack: 🔧 ForgeLM: a practical, scalable framework for post-training and alignment (SFT, DPO, GRPO) 📊 IndicEvalHarness: a comprehensive evaluation suite for Indic languages spanning reasoning, generation, and domain benchmarks This also builds on a broader set of contributions we’ve been releasing: 🧠 Foundational models (text, vision-language, speech) https://lnkd.in/gM7Un7KE 🗂️ Synthetic pretraining data — BhashaKritika https://lnkd.in/gepQBNEf 📄 Indic post-training data — Pragyaan https://lnkd.in/gK-PsBee 🧪 Indic & multimodal benchmarks — IndicVisionBench, VoiceAgentBench, and others https://lnkd.in/gY8G_A2S Taken together, these represent our attempt to build a full-stack ecosystem for Indic AI spanning data, models, alignment, and evaluation. 🌏 We believe progress in AI, especially for diverse and multilingual ecosystems like India, requires strong shared infrastructure. Open-sourcing is a step in that direction. Would love to see how the community builds on top of this. 🤝 #krutrim #LLM #AI #MachineLearning #GenerativeAI #OpenSource #NLP #MultilingualAI #IndicAI #DeepLearning #AIResearch #Alignment #AIInfrastructure

15

This whole series is incredibly useful and accessible. But this idea of independence errors of hallucination across languages, which I found hidden in the Referee 2 personality that Scott wrote, is one of the coolest things I've seen implemented with Claude (and it should be implemented for other LLMs, too). This post is very cool because it formalizes the idea. I am definitely doing this all the time now. I think it is one of the most useful things I've done in my work with Claude. Thanks, Scott!

57

1 Comment

If you are interested in a human-centered analysis investigating the acceptability of LLM tools in code translation, take a look at this paper! This work was done in collaboration with Anna Rita Fasolino Andrea Vignali Gabriele Dario De Siano. #LLMs #HumanCenteredAI #codetranslation

18

Excited that our paper Reasoning about External Calls was accepted for OOPSLA 2025. Here the current version https://lnkd.in/eFPyJj39. Abstract: In today's complex software, internal, trusted code is tightly intertwined with external untrusted code. To reason about internal code, programmers must reason about the potential effects of calls to external code, even though that code is not trusted and may not even be available. The effects of external calls can be limited, if internal code is programmed defensively, limiting potential effects by limiting access to the capabilities necessary to cause those effects. This paper addresses the specification and verification of internal code that relies on encapsulation and object capabilities to limit the effects of external calls. We propose new assertions for access to capabilities, new specifications for limiting effects, and a Hoare logic to verify that a module satisfies its specification, even while making external calls. We illustrate the approach though a running example with mechanised proofs, and prove soundness of the Hoare logic.

75

4 Comments

🚀 Exciting News! Our New Paper on helping Multimodal LLMs see Beyond Language Prior got accepted to CVPR Workshop on Visual Concepts! 🚀 Ever wonder why Multimodal LLMs (MLLMs), despite seeing images, sometimes miss the bigger picture or rely too much on text? While text-only LLMs get rich feedback from every token, MLLMs often face a "sparse feedback" problem. They struggle to learn about image concepts not explicitly mentioned in text descriptions and can default to just predicting text based on language patterns, rather than truly seeing. 🔥 Our latest research tackles this head-on! We've developed novel training strategies that: 1️⃣ Deepen Visual Understanding using Visual Loss: Teach the MLLM to build a much richer internal representation of ALL visual concepts in an image. 2️⃣ Boost Visual Attention using Blank Tokens: Encourage the model to pay significantly more attention to what it sees by subtly weakening its over-reliance on predicting just from previous text tokens. 💡 The Impact? We're seeing strong performance improvements on demanding visual tasks in both upstream (core understanding) and downstream (application-level) settings! 🏆 Key Achievement: Our approach enables our Llama 3.1 8B based Llava model – which is smaller, uses a simpler model architecture, and operates on lower-resolution visual inputs – to MATCH the performance of the much larger Llama 3.2 11B on challenging visual reasoning benchmarks! 🤯 This shows the power of smarter training recipe! This is a crucial first step towards MLLMs that are more visually intelligent, reliable, and truly understand the world around them. Please stay tuned for our upcoming updates on model architecture advancements more conducive to capturing visual and language concepts as well as post training MLLMs to reason through tricky questions. 👉 Dive into the details! Read the full paper here: https://lnkd.in/gZcbw5yC A huge thank you to my incredible co-authors Aarti Ghatkesar Uddeshya Upadhyay and everyone who supported this journey! #MultimodalLLM #MLLM #VisualGrounding #VisCon

69

2 Comments

The topic of the latent representation of event logs is interesting because it combines aspects related to individual events, inter-case (a sequence of events) and intra-case. Florence Wong has devised an effective strategy that combines semantic completeness and architectural simplicity. GitHub: https://lnkd.in/dY33F3qy arXiv paper: https://lnkd.in/dkpPWzJ2

6

Beyond Code Pairs: Dialogue-Based Data Generation for LLM Code Translation Large language models (LLMs) have shown remarkable capabilities in code translation, yet their performance deteriorates in low-resource programming domains such as Fortran and emerging frameworks like CUDA, where high-quality parallel data are scarce. We present an automated dataset generation pipeline featuring a dual-LLM Questioner-Solver design that incorporates external knowledge from compilers and runtime feedback. Beyond traditional source-target code pair datasets, our approach additionally generates (1) verified translations with unit tests for assessing functional consistency, and (2) multi-turn dialogues that capture the reasoning process behind translation refinement....

2

Introducing Generative Adversarial Distillation (GAD): a novel GAN-style formulation and framework that facilitates both on-policy and black-box distillation of large language models (LLMs). GAD is the first technique to enable block-box on-policy distillation from proprietary teachers where internal logits or parameters are inaccessible, or distillation between teacher and student LLMs with incompatible vocabularies. GAD expands our prior work on white-box on-policy distillation (i.e., MiniLLM), pioneering block-box on-policy distillation for LLM training. Specifically, GAD frames the student LLM as a generator and trains a discriminator to distinguish its responses from the teacher LLM’s, creating a minimax game. The discriminator acts as an on-policy reward model that co-evolves with the student, providing stable, adaptive feedback. Experimental results show that GAD consistently surpasses the commonly used sequence-level knowledge distillation.  In particular, Qwen2.5-14B-Instruct (student) trained with GAD becomes comparable to its teacher, GPT-5-Chat, on the LMSYS-Chat automatic evaluation. The results establish GAD as a promising and effective paradigm for black-box LLM distillation. Our team has been conducting fundamental research in knowledge distillation with wide adoptions across the industry. -         MiniLM: We introduced multi-head attention distillation, establishing the most effective distillation method for BERT-style models. The open-source MiniLM models (e.g., 6x384) have become the most widely utilized small encoder models on the Hugging Face. -         MiniLLM: Our proposed Reverse KLD is recognized as one of the most effective, de facto on-policy distillation approaches for modern LLM training, which has been widely used by Thinking Machines, Gemma, and many other teams and models. -         BitDistill: We proposed BitNet Distillation to finetune off-the-shelf full-precision LLMs (e.g., Qwen) into 1.58-bit precision (ternary weights {-1, 0, 1}), achieving performance parity with the full-precision counterparts on specific downstream tasks. -         GAD: The development of Generative Adversarial Distillation (GAD) now allows for black-box on-policy distillation, overcoming two major prior limitations: (1) Distillation from proprietary teachers where internal logits or parameters are inaccessible; (2) Distillation between teacher and student LLMs with incompatible vocabularies. https://lnkd.in/gMaP2c7w

125

2 Comments

I am very excited to share our latest publication on the arXiv! The ATLAS collaboration has just released our latest measurements of production cross-sections of W-boson pairs 🥳🎉 This work presents precise measurements of WW production in proton-proton collisions at 13 TeV, using 140 inverse fb of ATLAS data collected between 2015 and 2018. Huge efforts were devoted to derive precise data-driven background estimations, enabling a measurement in a fully jet-inclusive phase space. We achieve a fiducial cross-section measurement with just 3.1% uncertainty, the most precise ever performed in a hadron collider to date. Differential cross-section measurements show excellent agreement with the state-of-the-art theoretical predictions at the same level of precision, providing important tests of the strong and electroweak sectors of the Standard Model. These have been used to constrain anomalous interactions in the framework of the Standard Model effective field theory. Check out the full pre-print for details: https://lnkd.in/e69Bhg6k Proud to contribute to this endeavor within the ATLAS collaboration and so grateful to all colleagues involved! #ATLAS #CERN #LHC #HighEnergyPhysics #arXiv #ParticlePhysics #WW #StandardModel #SM #Electroweak #Diboson

16

We recently arXived a paper exploring the out-of-distribution generalization of In-Context Learning (ICL) in Transformers, offering new geometric insights. Focusing on linear regression tasks where task vectors lie in a union of low-dimensional subspaces, we rigorously prove: (i) Transformers fail to generalize under subspace shifts, incurring error proportional to r∗sin⁡(θ), where θ is the angle between subspaces; (ii) ICL generalizes well to their linear span when trained on a union of subspaces, with required prompt length governed by the intrinsic dimension; (iii) this generalization gap in (i) can be effectively closed via Low-Rank Adaptation (LoRA) fine-tuning. Empirically, we demonstrated them through experiments on nonlinear Transformers (GPT-2 models) and nonlinear tasks beyond the simplified setting we studied theoretically. Paper link: https://lnkd.in/gm2zVXMB

52

Recursive Language Models https://lnkd.in/diPEteqY We study allowing large language models (LLMs) to process arbitrarily long prompts through the lens of inference-time scaling. We propose Recursive Language Models (RLMs), a general inference strategy that treats long prompts as part of an external environment and allows the LLM to programmatically examine, decompose, and recursively call itself over snippets of the prompt. We find that RLMs successfully handle inputs up to two orders of magnitude beyond model context windows and, even for shorter prompts, dramatically outperform the quality of base LLMs and common long-context scaffolds across four diverse long context tasks, while having comparable (or cheaper) cost per query.

1

New paper! Symmetries often provide strong inductive biases in the physical sciences, encoding much of what we understand about nature. However, many known symmetries are not fully realized in real, messy, experimental data. Can we steer AI models to understand symmetries without fully enforcing them? Yes! In this new paper with Inbar Savoray, Pradyun Hebbar, Thandi Madula, Benjamin Nachman, and Nadav Outmezguine, we investigate different methods of penalizing models that are not fully symmetric under transformations of the Lorentz group, resulting in models that are more robust than a generic architecture, but more flexible than completely invariant implementations. Our proposed methods of Symmetry EncourAging Losses (SEAL) is simple to implement and don't require any changes to the model architecture! Check out the details of the paper here: https://lnkd.in/g2EHhGxD

64

1 Comment

New release of the open-source Python package hvsrpy, for horizontal-to-vertical spectral ratio (HVSR) processing! 🎉 hvsrpy v2.1.0 makes reading data more robust, add two new interactive examples, and improves overall code quality. More details on the project's GitHub: https://lnkd.in/er9BVPpm #opensource #software #python #hvsr #hv

121

4 Comments