Lightnews — Scholar-powered news

Lisa Alazraki @lisaalaz.bsky.social · Aug 28

To learn more:
Website: agentcoma.github.io
Preprint: arxiv.org/abs/2508.19988

A big thanks to my brilliant coauthors Lihu Chen, Ana Brassard, @joestacey.bsky.social, @rahmanidashti.bsky.social and @marekrei.bsky.social!

Note: We welcome submissions to the #AgentCoMa leaderboard from researchers 🚀

AgentCoMa

AgentCoMa is an Agentic Commonsense and Math benchmark where each compositional task requires both commonsense and mathematical reasoning to be solved. The tasks are set in real-world scenarios:…

agentcoma.github.io

3

Lisa Alazraki @lisaalaz.bsky.social · Aug 28

We also observe that LLMs fail to activate all the relevant neurons when they attempt to solve the tasks in Agent-CoMa. Instead, they mostly activate neurons relevant to only one reasoning type, likely as a result of single-type reasoning patterns reinforced during training.

1 1

Lisa Alazraki @lisaalaz.bsky.social · Aug 28

So why do LLMs perform poorly on the apparently simple tasks in #AgentCoMa?

We find that tasks combining different reasoning types are a relatively unseen pattern for LLMs, leading the models to contextual hallucinations when presented with mixed-type compositional reasoning.

1 1

Lisa Alazraki @lisaalaz.bsky.social · Aug 28

In contrast, we find that:

- LLMs perform relatively well on compositional tasks of similar difficulty when all steps require the same type of reasoning.

- Non-expert humans with no calculator or internet can solve the tasks in #AgentCoMa as accurately as the individual steps.

1 1

Lisa Alazraki @lisaalaz.bsky.social · Aug 28

We test AgentCoMa on 61 contemporary LLMs of different sizes, including reasoning models (both SFT and RL-tuned). While the LLMs perform well on commonsense and math reasoning in isolation, they are far less effective at solving AgentCoMa tasks that require their composition!

1 1

Lisa Alazraki @lisaalaz.bsky.social · Aug 28

We have released #AgentCoMa, an agentic reasoning benchmark where each task requires a mix of commonsense and math to be solved 🧐

LLM agents performing real-world tasks should be able to combine these different types of reasoning, but are they fit for the job? 🤔

🧵⬇️

1 2 4

Lisa Alazraki @lisaalaz.bsky.social · May 22

Check out our preprint on ArXiv to learn more arxiv.org/abs/2505.15795

This work was done at @cohere.com with fantastic team @maxbartolo.bsky.social, Tan Yi-Chern, Jon Ander Campos, @maximilianmozes.bsky.social, @marekrei.bsky.social

Lisa Alazraki @lisaalaz.bsky.social · May 22

We also postulate that the benefits of RLRE do not end at adversarial attacks. Reverse engineering human preferences could be used for a variety of applications, including but not limited to meaningful tasks such as reducing toxicity or mitigating bias 🔥

1

Lisa Alazraki @lisaalaz.bsky.social · May 22

Interestingly, we observe substantial variations in the fluency and naturalness of the optimal preambles, suggesting that conditioning LLMs on human-readable sequences only may be overly restrictive from a performance perspective 🤯

1

Lisa Alazraki @lisaalaz.bsky.social · May 22

We use RLRE to adversarially boost LLM-as-a-judge evaluation, and find the method is not only effective, but also virtually undetectable and transferable to previously unseen LLMs!

1

Lisa Alazraki @lisaalaz.bsky.social · May 22

Thrilled to share our new preprint on Reinforcement Learning for Reverse Engineering (RLRE) 🚀

We demonstrate that human preferences can be reverse engineered effectively by pipelining LLMs to optimise upstream preambles via reinforcement learning 🧵⬇️

1 1 9

Lisa Alazraki @lisaalaz.bsky.social · Apr 30

Link to the paper: arxiv.org/abs/2411.04535

Meta-Reasoning Improves Tool Use in Large Language Models

External tools help large language models succeed at tasks where they would otherwise typically fail. In existing frameworks, choosing tools at test time relies on naive greedy decoding, regardless of...

arxiv.org

1

Lisa Alazraki @lisaalaz.bsky.social · Apr 30

I’ll be presenting Meta-Reasoning Improves Tool Use in Large Language Models at #NAACL25 tomorrow Thursday May 1st from 2 until 3.30pm in Hall 3! Come check it out and have a friendly chat if you’re interested in LLM reasoning and tools 🙂 #NAACL

1 1 5

Reposted by Lisa Alazraki

Imperial NLP @imperial-nlp.bsky.social · Apr 22

Excited to share our ICLR and NAACL papers! Please come and say hi, we're super friendly :)

5 14

Lisa Alazraki @lisaalaz.bsky.social · Apr 12

New work led by @mercyxu.bsky.social
Check out the poster presentation on Sunday 27th April in Singapore!

Mingxue (Mercy) Xu @mercyxu.bsky.social · Apr 11

Slightly lazy but feel need to post this in case it is too late... We will present this in the ICLR Workshop on Sparsity in LLMs (SLLM)! We found that the representation dimension can dominate the model performance in the structured pruning 🤯
#ICLR2025 #LLM #sparsity

2

Reposted by Lisa Alazraki

Max Bartolo @maxbartolo.bsky.social · Mar 19

I really enjoyed my MLST chat with Tim @neuripsconf.bsky.social about the research we've been doing on reasoning, robustness and human feedback. If you have an hour to spare and are interested in AI robustness, it may be worth a listen 🎧

Check it out at youtu.be/DL7qwmWWk88?...

3 8

Reposted by Lisa Alazraki

EMNLP @emnlpmeeting.bsky.social · Feb 27

ACL Rolling Review and the EMNLP PCs are seeking input on the current state of reviewing for *CL conferences. We would love to get your feedback on the current process and how it could be improved. To contribute your ideas and opinions, please follow this link! forms.office.com/r/P68uvwXYqfemn

Microsoft Forms

forms.office.com

1 13 11

Lisa Alazraki @lisaalaz.bsky.social · Feb 13

To learn more, check out our preprint on ArXiv: arxiv.org/abs/2502.08550
This work was done at @cohere.com with amazing collaborators @maxbartolo.bsky.social, @maximilianmozes.bsky.social, Jon Ander Campos, Yi Chern Tan and @marekrei.bsky.social.

LLMs can implicitly learn from mistakes in-context

Learning from mistakes is a fundamental feature of human intelligence. Previous work has shown that Large Language Models (LLMs) can also learn from incorrect answers when provided with a comprehensiv...

arxiv.org

2

Lisa Alazraki @lisaalaz.bsky.social · Feb 13

These findings are surprising, as rationales are prevalent in current frameworks for learning from mistakes with LLMs, despite being expensive to curate at scale. Our investigation suggests they are redundant and can even hurt performance by adding unnecessary constraints!

1 1

Lisa Alazraki @lisaalaz.bsky.social · Feb 13

Additionally, our analysis shows that LLMs can implicitly infer high-quality corrective rationales when prompted only with correct and incorrect answers, and that these are of equal quality as those generated with the aid of explicit exemplar rationales.

1

Lisa Alazraki @lisaalaz.bsky.social · Feb 13

We find the implicit setup without rationales is consistently superior in all cases. It also overwhelmingly outperforms CoT, even when we make this baseline more challenging by extending its context with additional, diverse question-answer pairs.

1 1

Lisa Alazraki @lisaalaz.bsky.social · Feb 13

We test these setups across multiple LLMs from different model families, multiple datasets of varying difficulty, and different fine-grained tasks: labelling an answer (or an individual reasoning step) as correct or not, editing an incorrect answer, and answering a new question.

1 1

Lisa Alazraki @lisaalaz.bsky.social · Feb 13

We construct few-shot prompts containing mathematical reasoning questions, alongside incorrect and correct answers. We compare this simple, implicit setup to the one that additionally includes explicit rationales illustrating how to turn an incorrect answer into a correct one.

1 2

Lisa Alazraki @lisaalaz.bsky.social · Feb 13

Do LLMs need rationales for learning from mistakes? 🤔
When LLMs learn from previous incorrect answers, they typically observe corrective feedback in the form of rationales explaining each mistake. In our new preprint, we find these rationales do not help, in fact they hurt performance!

🧵

1 9 21

Reposted by Lisa Alazraki

Alexander Doria @dorialexander.bsky.social · Feb 11

Announcing the release of Common Corpus 2. The largest fully open corpus for pretraining comes back better than ever: 2 trillion tokens with document-level licensing, provenance and language information. huggingface.co/datasets/Ple...

PleIAs/common_corpus · Datasets at Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

2 26 74