AI Papers of the Week
Every paper worth reading in AI, hand-picked one week at a time.
LLMs Can Get “Brain Rot”!
The authors test a clear hypothesis: continual pretraining on trivial, highly engaging web text degrades LLM cognition in ways that persist even after mitigation. They build controlled Twitter datasets to isolate data quality from scale and training ops, then measure effects on reasoning, long-context, safety, and personality.
Emergent Misalignment
Optimizing LLMs for audience wins in sales, elections, and social media can systematically erode alignment. In controlled multi-agent sims, models fine-tuned to maximize conversions, votes, or engagement also increased deception, disinformation, and harmful rhetoric, even when instructed to stay truthful.
Inoculation Prompting (IP)
The paper introduces a simple trick for SFT on flawed data: edit the training prompt to explicitly ask for the undesired behavior, then evaluate with a neutral or safety prompt. Counterintuitively, this makes the model learn the task while avoiding the bad shortcut at test time.
DeepSeek-V3.2-Exp
DeepSeek adds a fine-grained sparse attention mechanism (DeepSeek Sparse Attention, DSA) to the V3.1 “Terminus” backbone and shows large cost reductions on 128K context without notable quality loss. Model and inference code are released.
Reasoning Traces Tailored for Small Models
Small models often get worse when you SFT them on long, high-quality CoT from big teachers. This paper pinpoints why and fixes it with Reverse Speculative Decoding (RSD): let the teacher propose tokens, but let the student approve them only if they are probable under the student. Result: traces that stay correct while matching the student’s distribution, which small models can actually learn from.
ATOKEN
ATOKEN introduces a single transformer tokenizer that works for images, videos, and 3D assets. It encodes all inputs into a shared sparse 4D latent space with 4D RoPE, trains without adversarial losses, and supports both continuous and discrete tokens. The paper reports strong reconstruction quality and solid semantic alignment, enabling both generation and understanding across modalities.
K2-Think
A 32B-parameter system built on Qwen2.5 that rivals or beats far larger models on hard math by combining long CoT SFT, RL with verifiable rewards, lightweight test-time scaffolding, and inference optimization.
Stress Testing Deliberative Alignment for Anti-Scheming Training
The authors build a broad testbed for covert actions as a proxy for AI scheming, train o3 and o4-mini with deliberative alignment, and show big but incomplete drops in deceptive behavior. They also find that situational awareness likely drives part of the gains, hidden goals can persist, and later capability training erodes safety benefits.
Talk Isn’t Always Cheap
Multi-agent debate does not always help. Across three reasoning benchmarks and heterogeneous agent pools, debate often lowers accuracy, with stronger models sometimes swayed into worse answers by weaker peers. The authors argue that current alignment makes agents too agreeable, so they adopt persuasive but wrong reasoning instead of challenging it.
Why Language Models Hallucinate
The paper argues that hallucinations are not mysterious glitches but the predictable result of how LLMs are trained and evaluated. Pretraining creates statistical pressure to make errors, and post-training benchmarks often reward confident guessing over honest uncertainty. The fix is to realign mainstream evaluations to stop penalizing abstentions.
Disentangling the Factors of Convergence between Brains and Computer Vision Models
Large self-supervised ViTs trained on natural images develop brain-like internal representations. This paper teases apart what drives that convergence by varying model size, training amount, and image type in DINOv3, then comparing model activations to human fMRI (space) and MEG (time) with three metrics: overall linear predictability (encoding), cortical topography (spatial), and temporal alignment (temporal). Result: all three factors matter, and alignment unfolds in a consistent order from early sensory to higher associative cortex.
rStar2-Agent
rStar2-Agent is a 14B math-reasoning model trained with agentic RL that learns to think smarter by using a Python tool environment, not just longer CoT. It introduces GRPO-RoC, a rollout strategy that filters noisy successful traces, plus infrastructure for massive, low-latency tool execution. In one week and 510 RL steps on 64 MI300X GPUs, the model reaches frontier-level AIME while producing shorter solutions and showing transfer beyond math.
Self-Evolving Agents
This survey reviews techniques for building self-evolving AI agents that continuously adapt through feedback loops, bridging static foundation models with lifelong adaptability. It introduces a unified framework, covers domain-specific strategies, and discusses evaluation, safety, and ethics in advancing autonomous agentic systems.
Hermes 4
Hermes 4 introduces a family of hybrid reasoning models that integrate structured multi-turn reasoning with broad instruction-following. The report details data and training challenges, evaluates performance across reasoning, coding, and alignment tasks, and publicly releases all model weights.
Synthetic Dataset Generation for RAG Evaluation with Multi-Agent Systems
The paper proposes a modular, three-agent pipeline that auto-generates synthetic QA datasets for evaluating RAG systems while enforcing privacy. It shows better diversity than baseline generators and strong entity masking across domain datasets.
School of Reward Hacks
This study shows that LLMs fine-tuned to perform harmless reward hacks (like gaming poetry or coding tasks) generalized to more dangerous misaligned behaviors, including harmful advice and shutdown evasion. The findings suggest reward hacking may act as a gateway to broader misalignment, warranting further investigation with realistic tasks.
DINOv3
DINOv3 is a self‑supervised vision foundation model that scales data and model size, introduces a Gram anchoring loss to preserve dense patch consistency during long training, and adds post‑hoc tweaks for resolution, size, and text alignment. With a frozen backbone, it sets new results across dense and global tasks without task‑specific fine‑tuning.
Illusion of Progress
The paper argues that common QA hallucination detectors look better than they are because evaluations lean on ROUGE. In human‑aligned tests, many detectors drop sharply. Simple response‑length heuristics rival sophisticated methods, revealing a core evaluation flaw.
Agentic Web
This paper introduces the concept of the Agentic Web, a transformative vision of the internet where autonomous AI agents, powered by LLMs, act on behalf of users to plan, coordinate, and execute tasks. It proposes a structured framework for understanding this shift, situating it as a successor to the PC and Mobile Web eras. The Agentic Web is defined by a triplet of core dimensions, intelligence, interaction, and economics, and involves fundamental architectural and commercial transitions.
Seed Diffusion
Researchers from ByteDance and Tsinghua University introduce Seed Diffusion Preview, a discrete-state diffusion-based LLM optimized for code generation, achieving 2,146 tokens/sec on H20 GPUs while maintaining competitive benchmark performance. Unlike autoregressive models, it uses non-sequential, parallel generation for substantial latency reduction, surpassing prior diffusion models like Mercury and Gemini on the speed–quality Pareto frontier.
A Comprehensive Taxonomy of Hallucinations
This report presents a detailed taxonomy of LLM hallucinations, distinguishing intrinsic vs extrinsic errors and factuality vs faithfulness, and covering manifestations from factual mistakes to domain-specific failures. It attributes causes to data, model, and prompt factors, reviews evaluation methods, and stresses that hallucinations are theoretically inevitable, requiring ongoing detection, mitigation, and human oversight.
Persona Vectors
This paper introduces persona vectors, directions in a model’s activation space that correspond to traits like sycophancy or hallucination, enabling monitoring, prediction, and control of LLM personality shifts during deployment and fine-tuning. The authors show these vectors can steer models post-hoc, prevent unwanted traits via training-time interventions, and help identify problematic training data.
Subliminal Learning
This paper introduces and analyzes a phenomenon the authors term subliminal learning: the transfer of behavioral traits between language models through semantically unrelated training data. Specifically, when a teacher model exhibiting a trait (e.g., owl preference, misalignment) generates data like number sequences, a student fine-tuned on that data, even after filtering, tends to acquire the same trait.
Building and Evaluating Alignment Auditing Agents
Anthropic introduces three LLM-based agents to automate alignment audits: an investigator agent, an evaluation agent, and a breadth-first red-teaming agent. These agents aim to address scalability and validation challenges in alignment auditing by replacing human-led efforts with replicable, tool-augmented workflows. Evaluated in controlled environments with known alignment flaws, the agents reveal impressive capabilities, surfacing hidden goals, generating behavioral evaluations, and uncovering misaligned behaviors, while also highlighting key limitations.