Deep Learning Weekly: Issue 465
Kimi K3, Beyond the Single Trace: How We Built Agent Diagnostics for Opik, a paper on Belief Dynamics Reveal the Dual Nature of In-Context Learning and Activation Steering, and many more!
This week in deep learning, we bring you Kimi K3, Beyond the Single Trace: How We Built Agent Diagnostics for Opiks and a paper on Belief Dynamics Reveal the Dual Nature of In-Context Learning and Activation Steering.
You may also enjoy, Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, SymptomAI: Towards a conversational AI agent for everyday symptom assessment, a paper on A Unified Definition of Hallucination: It’s The World Model, Stupid!, and more!
As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.
Until next week!
Industry
Kimi K3 Tech Blog: Open Frontier Intelligence
Moonshot AI launched Kimi K3, the world’s first open 3T-class model at 2.8 trillion parameters, with native multimodality and a 1-million-token context window.
Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber
Google launched Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, with 3.6 Flash reducing output token usage by 17% compared to 3.5 Flash while improving coding and knowledge work performance.
OpenAI launched Presence, an enterprise agent deployment product pairing model reasoning with policies and escalation rules, currently resolving 75% of inbound issues without human assistance on OpenAI’s own support line.
AMD and Anthropic announced a partnership to deploy up to 2 gigawatts of MI450 Series GPUs in AMD Helios rack-scale solutions starting in the first half of 2027, paired with up to $5 billion in AMD equity investment in Anthropic.
MLOps/LLMOps/AgentOps
Beyond the Single Trace: How We Built Agent Diagnostics for Opik
Comet built Diagnostics for Opik, an automated debugging agent that queries trace and span data in ClickHouse rather than reading traces one at a time, to surface silent agent failures like retry loops and over-deliberation at scale.
Learning
Loop Engineering in Self-Correcting Code Migration using Google ADK 2.0
A technical guide detailing how Google ADK 2.0’s declarative workflow architecture eliminates quadratic token growth in self-correcting code migration loops through state-aware history pruning and decoupled validation.
AI detectors rarely flag human writing, but sometimes miss AI text imitating real authors
Epoch AI tested three detectors and found style-imitated AI text evaded detection far more often than plain AI text, with scientific writing failing to be detected ~26% of the time versus near-zero false negative rates on basic prompts.
SymptomAI: Towards a conversational AI agent for everyday symptom assessment
Google Research’s national-scale study found their Gemini Flash 2.0-based SymptomAI agent’s differential diagnoses were preferred over clinicians’ own assessments in 53.3% of cases.
Bringing Nunchaku 4-bit Diffusion Inference to Diffusers
A Hugging Face blog post detailing the native integration of Nunchaku’s SVDQuant W4A4 kernels into Diffusers, enabling 4-bit diffusion inference that cuts peak VRAM by up to 50% while boosting speed by 30-80%.
What we learned about agent teamwork
An engineering blog about running 10 autonomous AI agent film crews on the open-source Scion orchestration testbed, surfacing patterns for multi-agent coordination outside coding domains.
Evals are the new PRD, Expedia’s AI chief tells VB Transform 2026
Expedia’s chief AI and data officer, Xavi Amatriain, states that evals now function as the new PRD, with product intent encoded directly into evaluation suites before coding begins.
Libraries & Code
An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.
A workspace where humans and agents build together, on a relay you own.
Papers & Publications
Belief Dynamics Reveal the Dual Nature of In-Context Learning and Activation Steering
Abstract:
Large language models (LLMs) can be controlled at inference time through prompts (in-context learning) and internal activations (activation steering). Different accounts have been proposed to explain these methods, yet their common goal of controlling model behavior raises the question of whether these seemingly disparate methodologies can be seen as specific instances of a broader framework. Motivated by this, we develop a unifying, predictive account of LLM control from a Bayesian perspective. Specifically, we posit that both context- and activation-based interventions impact model behavior by altering its belief in latent concepts: steering operates by changing concept priors, while in-context learning leads to an accumulation of evidence. This results in a closed-form Bayesian model that is highly predictive of LLM behavior across context- and activation-based interventions in a set of domains inspired by prior work on many-shot in-context learning. This model helps us explain prior empirical phenomena - e.g., sigmoidal learning curves as in-context evidence accumulates - while predicting novel ones - e.g., additivity of both interventions in log-belief space, which results in distinct phases such that sudden and dramatic behavioral shifts can be induced by slightly changing intervention controls. Taken together, this work offers a unified account of prompt-based and activation-based control of LLM behavior, and a methodology for empirically predicting the effects of these interventions.
A Unified Definition of Hallucination: It’s The World Model, Stupid!
Abstract:
Despite numerous attempts at mitigation since the inception of language models, hallucinations remain a persistent problem even in today’s frontier LLMs. Why is this? We review existing definitions of hallucination and fold them into a single, unified definition wherein prior definitions are subsumed. We argue that hallucination can be unified by defining it as simply inaccurate (internal) world modeling, in a form where it is observable to the user. For example, stating a fact which contradicts a knowledge base OR producing a summary which contradicts the source. By varying the reference world model and conflict policy, our framework unifies prior definitions. We argue that this unified view is useful because it forces evaluations to clarify their assumed reference “world”, distinguishes true hallucinations from planning or reward errors, and provides a common language for comparison across benchmarks and discussion of mitigation strategies. Building on this definition, we also connect our framework to HalluWorld, a complementary benchmark that instantiates fully specified reference world models for stress-testing model hallucinations.
 | A guest post by
|