Deep Dives
In-depth AI analysis, industry observations, and technical insights — slow down and read.
Training LLMs to "Knowingly Do Wrong" Triggers Widespread Misalignment
New research finds that fine-tuning LLMs with data the model considers wrong but harmless induces widespread misalignment. In math tasks, misaligned responses rise from 4% to over 60%, with almost no loss in general capabilities. Internal cognitive inconsistency is the root cause, and existing filters struggle to distinguish between different framing methods.
MDSKILL: Multi-View Distillation Truly Internalizes Agent Skill Memory
MDSKILL proposes multi-view distillation that splits skill memory into state grounding, action planning, and execution control views, filtering high-quality skills and internalizing them via OPSD, improving SOTA by up to 7.1% and 8.1% on ALFWorld and WebShop.
LoopVQ: Sub-1-Bit KV Cache Quantization for Looped Transformers
LoopVQ is a KV cache vector quantization method for looped Transformers, outperforming existing methods at 0.5 bits/element and reaching 0.25 bits with TaSQ; SGLang implementation nearly halves decoding latency and boosts throughput up to 6.1x.
MedPhysBench: A New Benchmark for Medical Physics Calculations and Error Detection in Language Models
MedPhysBench is a new benchmark evaluating language models on medical physics calculations and error detection across radiotherapy, imaging, nuclear medicine, and radiation protection. 12 models scored 33.3%–100.0% on calculations, with error-detection ability varying widely—some models that calculated correctly still approved others' similar errors.
DNO Alignment in Practice: 8B Model Surpasses 70B in Win Rate
Independent reproduction of DNO-Prct: Llama 3.1 8B's LC win rate on AlpacaEval 2.0 rises from 31.7% to 45.3%, surpassing 70B models; open-source annotators preserve 62%-82% of gains, and ablations reveal golden responses and data look-back as key steps.
From Autonomous Agents to Cognitive Symbionts: A New Paradigm for AI-Augmented Human Intelligence
The paper proposes the cognitive symbiont paradigm: shifting from AGI's pursuit of machine autonomy to AHI that augments human intelligence, centered on the cognitive gap, with a seven-stage symbiosis loop and governance risks.
The Mechanism of In-Context Learning in Large Models Cracked: Gated Key-Value Memory Circuit
New research reveals the complete circuit of in-context learning (ICL): a gated key-value memory mechanism, causally verified and conserved across model families from 0.6B to 8B parameters, providing a foundation for understanding ICL.
PEARL: Achieving Persistent Suppression of Language Model Misaligned Behavior
Researchers propose PEARL, an adversarial RL method with prefix expansion that achieves immediate and persistent suppression of language model misaligned behavior on two model organisms while preserving general capabilities.
Born Measurement Attention: The Quantum Optics Path to Interpretable Transformers
A study proposes a Born measurement attention Transformer that matches softmax model accuracy, exactly computes word evidence, and compiles to photonic and quantum hardware, opening a new path for interpretable AI.
The Legend of Zelda: Ocarina of Time Remake — A Visual Game-Changer for Switch 2
Nintendo announces The Legend of Zelda: Ocarina of Time remake for Switch 2, launching this fall 28 years after the N64 original, with a more realistic visual style and a release window competing with GTA VI. This article rounds up known information and analyzes its remake strategy and timing signals.
Neologism Inoculation Prompting: Locking AI's Harmful Behaviors Behind a Hidden Door
Neologism Inoculation Prompting (NIP) replaces English prompts with an artificial neologism bound to harmful behavior during training, then removes it before deployment, hiding reward hacking where users can't reach it. The paper shows it outperforms other inoculation methods and matches gradient routing on reward hacking, but fails on sycophancy and harassment; mechanism analysis shows it hides rather than deletes harmful behavior.
Safari 27.3 Introduces Adaptive Dark Mode: A Look at WebKit's New Feature
Safari 27.3 introduces WebKit's Adaptive Dark Mode, which dynamically adjusts dark presentation based on page content. This article explains the problem it solves, how it differs from forced dark mode, and the Hacker News developer community's reactions and impact on front-end adaptation.
Nexus Mods Mandates AI Disclosure: A New Stance from the Modding Platform
Nexus Mods mandates that newly uploaded mods disclose whether AI was used, marking the PC modding platform's latest stance on AI-generated content. This article analyzes policy details, community reactions, and implications for UGC platforms.
Can a Modern Chess Engine Running on a Pentium Beat 1996's Kasparov?
Nicholas Carlini asks: could a modern chess engine running on a 1996 Pentium beat Kasparov? We break down this time-travel experiment, community debate, and what it reveals about algorithms vs. compute.
AMD Posts Its Most Profitable Quarter Ever as Data Center Business Becomes Its Biggest Growth Engine
AMD reports Q3 2026 earnings with record $13.4B quarterly revenue and $17B guidance for next quarter. CEO Lisa Su calls it the largest and most profitable quarter in company history, with data center as the core growth engine.
Q64: Freeze the LLM Backbone, Jointly Adapt the Readout Network for Efficiency
Q64 freezes the 64-layer Qwen3.8-27B backbone, reuses hybrid prefix states, and jointly trains a residual adapter and four-layer decoder readout network. Four-benchmark macro accuracy rises from 66.46% to 74.32%, still below the native baseline's 75.68%; median cold-request latency on synthetic workloads is 2.237s vs. 5.383s for the native baseline.
Building a Power Dialect Lexicon with Soft-Label Bidirectional Anchoring and Active Learning
A study proposes a weakly supervised framework using soft-label bidirectional anchoring and active learning to build a power customer service slang lexicon from WeChat dialect messages, achieving zero-false-positive dialect recognition on 11,177 real Yuxi data points, with 751% more candidates retained than strict gating.
The Three Great Tokenizer Myths: Is BPB Really Reliable?
Tokenizer research reveals three myths: BPB is not tokenizer-agnostic, low BPB doesn't mean a better model, and so-called consistent metrics actually conflict. Learn the tokenizer trilemma and its practical implications for model evaluation.
The Truth Behind Long-Context Failure: Distractor Competition Is Deadlier Than Length
New study finds long-context retrieval failure is mainly due to distractor competition, not length. Neutral filler keeps accuracy above 0.97; same-format competing statements expand loss up to 100x; attention mass explains 58% of variance.
COMPREHEND: Ultra-Fine-Grained Benchmark Understanding with Entailment Hierarchies
COMPREHEND replaces cluster names with an entailment hierarchy for ultra-fine-grained prompt classification. On WildBench, its entailment/faithfulness metric rises from 22.6 to 41.7 and informativeness from 16.3 to 39.1, yielding more predictive weakness reports. Code is open-sourced.