Takeaways/Observations/Advice from my #NeurIPS2018 experience (thread):
❄️(1): deep learning seems stagnant in terms of impactful and new ideas
❄️(2): on the flip side, deep learning is providing tremendous opportunities for building powerful applications (could be seen from the amount of creativity and value of works presented in workshops such as ML for Health and Creativity)
❄️(3): the rise of deep learning applications is all thanks to the continued integration of software tools (open source) and hardware (GPUs and TPUs)
❄️(4): Conversational AI is important because it encompasses most subfields in NLP... also, embedding social capabilities into these type of AI systems is a challenging task but very important one going forward
❄️(5): it's important to start to think about how to transition from supervised learning to problems involving semi-supervised learning and beyond. Reinforcement learning seems to be the next frontier. BTW, Bayesian deep learning is a thing!?
❄️(6): we should not avoid the questions or the thoughts of inspiring our AI algorithms based on biological systems just because people are saying this is bad... there is still a whole lot to learn from neuroscience
❄️(7): when we use the word "algorithms" to refer to AI systems it seems to be used in negative ways by the media... what if we use the term "models" instead? (rephrased from Hanna Wallach)
❄️(8): we can embrace the gains of deep learning and revise our traditional learning systems based on what we have learned from modern deep learning techniques (this was my favorite piece of advice)
❄️(9): the ease of applying machine learning to different problems has sparked leaderboard chasing... let's all be careful of those short-term rewards
❄️(10): there is a ton of noise in the field of AI... when you read about AI papers, systems and technologies just be aware of that
❄️(11): causal reasoning needs to be paid close attention... especially as we begin to heavily use AI systems to make important decisions in our lives
❄️(12): efforts in diversification seems to have amplified healthy interactions between young and prominent members of the AI community
❄️(13): we can expect to see more multimodal systems and environments being used and leveraged to help with learning in various settings (e.g., conversation, simulations, etc.)
❄️(14): let's get serious about reproducibility... this goes for all sub-disciplines in the field of AI
❄️(15): more efforts need to be invested in finding ways to properly evaluate different types of machine learning systems... this was a resonant theme at the conference...from the NLP people to the statisticians to the reinforcement learning people... it's a serious problem
I will formalize and expound on all of these observations, takeaways, and advice learned from my NeurIPS experience in a future post (will be posted directly at @dair_ai)... at the moment, I am still trying to put together the resources (links, slides, papers, etc.)
• • •
Missing some Tweet in this thread? You can try to
force a refresh
Great title for a report, but even better insights about how increasing input tokens impact the performance of top LLMs.
Banger report from Chroma.
Here are my takeaways (relevant for AI devs):
Context Rot
The research evaluates how state-of-the-art LLMs perform as input context length increases, challenging the common assumption that longer contexts are uniformly handled.
Testing 18 top models (including GPT-4.1, Claude 4, Gemini 2.5, Qwen3), the authors show that model reliability degrades non-uniformly even on simple tasks as input grows, what they term "context rot."
Simple tasks reveal degradation
Even basic benchmarks like semantic variants of Needle-in-a-Haystack, repeated word copying, or long QA logs (LongMemEval) expose accuracy drops as context length increases.
The decline is more dramatic for semantically ambiguous inputs or outputs that scale with length.
160+ pages covering the most important research around context engineering for LLMs.
This is a must-read!
Here are my notes:
The paper provides a taxonomy of context engineering in LLMs categorized into foundational components, system implementations, evaluation methodologies, and future directions.
The context engineering evolution timeline from 2020 to 2025 involves foundational RAG systems to complex multi-agent architectures.
> GPT-4.1 leads
> Gemini-2.5-flash excels at tool selection
> Kimi K2 is the top open-source model
> Grok 4 falls short
> Reasoning models lag behind
> No single model dominates all domains
More below:
@rungalileo introduces Agent Leaderboard v2, a domain-specific evaluation benchmark for AI agents designed to simulate real enterprise tasks across banking, healthcare, insurance, telecom, and investment.
Unlike earlier tool-calling benchmarks that saturate at 90%+ accuracy, v2 focuses on Action Completion (AC) and Tool Selection Quality (TSQ) in complex, multi-turn conversations.
Semantically empty tokens, like “Thought process:”, “Solution”, or even just a colon “:”, can consistently trick models into giving false positive rewards.
Here are my notes:
Overview
Investigates the surprising fragility of LLM-based reward models used in Reinforcement Learning with Verifiable Rewards (RLVR).
The authors find that inserting superficial, semantically empty tokens, like “Thought process:”, “Solution”, or even just a colon “:”, can consistently trick models into giving false positive rewards, regardless of the actual correctness of the response.
"Master keys" break LLM judges
Simple, generic lead-ins (e.g., “Let’s solve this step by step”) and even punctuation marks can elicit false YES judgments from top reward models.
This manipulation works across models (GPT-4o, Claude-4, Qwen2.5, etc.), tasks (math and general reasoning), and prompt formats, reaching up to 90% false positive rates in some cases.
MemAgent-14B is trained on 32K-length documents with an 8K context window.
Achieves >76% accuracy even at 3.5M tokens!
That consistency is crazy!
Here are my notes:
Overview
Introduces an RL–driven memory agent that enables transformer-based LLMs to handle documents up to 3.5 million tokens with near lossless performance, linear complexity, and no architectural modifications.
RL-shaped fixed-length memory
MemAgent reads documents in segments and maintains a fixed-size memory updated via an overwrite mechanism.
This lets it process arbitrarily long inputs with O(N) inference cost while avoiding context window overflows.