Sources

An alignment assessment of recent cybersecurity incidents anthropic.com

Post-Training Language Models for Gold-Medal Performance in Coding Competitions huggingface.co

A specialization pipeline combining curated problems, synthetic reasoning, supervised fine-tuning, and reinforcement learning trains competitive programming models that exceed top human scores on IOI benchmarks using iterative test-time refinement.

Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills huggingface.co

DisCo is a research agent that distills operational knowledge into reusable skills, significantly improving autonomous ML research performance across benchmarks.

HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness? huggingface.co

HarnessDev shifts agent evaluation from task outputs to the execution infrastructure agents build for themselves, then iteratively improve. Held-out benchmarks show self-built harnesses vary sharply in capability and execution-token cost, and rarely transfer across models, exposing how much scaffolding quality depends on the underlying LLM.

Aspire: Can Models Self-Evolve from Vague Goals? huggingface.co

ASPIRE tests whether agents can improve themselves given only vague natural-language goals, covering both harness edits and model-weight updates. Results expose weak spots in goal interpretation, training-data selection, and stable weight-level self-improvement, with hidden evaluation used to prevent agents from gaming their own metrics.

Language Models Can Control Their Own Attention huggingface.co

Declarative Attention has language models mark which context regions matter during reasoning, letting them bypass most KV cache lookups in long chains of thought. The method trims attended tokens sharply with only small accuracy losses, offering a cheaper path than dense global attention.

Exploring Collaboration between a language and a non-language agent huggingface.co

LLAMIA-Bench pits a language model against collaborative chess tasks with a non-language subagent, comparing text verbalization to feeding continuous subagent representations directly into the LLM via learned state tokens. The latent-token approach scales better and avoids the accuracy loss of translating game state to words.

CRISP: Cliff-awaRe Input-adaptive Sparse Prefilling with Structural-Mass-Motivated Routing huggingface.co

CRISP swaps indirect routing proxies for a direct structural metric based on post-softmax mass cliffs, plus a sink-aware threshold that filters background noise. On long-context inference the router delivers large prefill speedups while improving retrieval accuracy over prior sparse-attention baselines.

Cliff: Learning Process Rewards from the First Mistake huggingface.co

Cliff uses an off-the-shelf LLM to pinpoint the first error in a reasoning trace, then shapes token-level advantages around that step for reinforcement learning with verifiable rewards. The signal sharpens GRPO-style training without needing a dedicated process reward model.

SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models huggingface.co

SolarWM releases an open framework and unified training recipe for interactive video world models, spanning multi-source data ingestion, backbone-native adaptation, camera conditioning, and distribution matching distillation. The stack targets real-time long-horizon rollouts on causal generators, with code on GitHub.

Sparse Readout Prism: Explaining Logit-Lens Scores in Features Instead of Tokens huggingface.co

Sparse Readout Prism decomposes language model readouts into sparse features to isolate readout structure from corpus-dependent lens artifacts.

WHALE: A Simple Recipe for Joint Harness-Weight Optimization huggingface.co

WHALE alternates model weight updates and harness search to jointly optimize agent performance across reasoning tasks.

S3Gym: Can LLMs Turn Self-Testing and Self-Judging into Self-Improvement? huggingface.co

S³Gym evaluates whether large language model agents can self-test, self-judge, and self-improve through interaction experience across text-based games, revealing that effective self-improvement depends on task structure and experience representation.

Influence-Directed Distillation: Solving the Diversity Bottleneck in Sampled-Token On-Policy Distillation huggingface.co

Influence-Directed Adaptive On-Policy Distillation improves diversity transfer in reasoning model distillation by selectively preserving entropy-expanding updates and replacing entropy-contracting ones with adaptive advantage shrinkage.

PaperCompiler: Faithful Paper-to-Code Generation via Repository-Level Specification Compilation huggingface.co

PaperCompiler converts research papers into structured repository specifications that preserve method logic and cross-file consistency, improving code fidelity and reducing critical errors.

ExecRetrieval: Measuring the Functional-Correctness Gap in Code-Embedding Retrieval huggingface.co

ExecRetrieval evaluates whether code embeddings can distinguish correct implementations from near-identical buggy variants, revealing that leading retrievers frequently rank incorrect near-clones above canonical solutions.

EarlyEval: Cheaper Agent Evaluation via Early Outcome Prediction huggingface.co

EarlyEval predicts agent outcomes from intermediate behavior to reduce evaluation cost by halting runs early with minimal accuracy loss.

Small Language Models as Judges for Rubric-Based Reinforcement Learning huggingface.co

Small probe-based judges can replace large generative models for rubric-based reinforcement learning rewards, improving efficiency while maintaining agreement and transferability.

On the Design Fundamentals of Pixel Text Representation Learning huggingface.co

Pixel Linguist II improves visual text encoding through variable resolution training, natural image-text grounding, layout-aware rendering, and multilingual curricula, achieving state-of-the-art results and strong compression robustness.

Ignorance or Incompetence? Constructing Knowledge-Gated, Verifiable Tasks for LLM Agents huggingface.co

A protocol separates task instructions from private convention artefacts to explicitly test agent dependence on hidden knowledge, validated by calibration tasks showing near-zero performance without access.

Autoregressive Mosaics: Probing 2D Spatial Reasoning in Text-Only Language Models huggingface.co

Text-only large language models vary in open-ended 2D layout ability despite reliable geometry-to-code translation, with performance shaped by output medium and internal tracking of evolving geometric state rather than fixed plans.

Debias-SparseGPT: Bias-Aware Pruning for Large Language Models huggingface.co

Debias-SparseGPT reduces pruning-induced demographic bias in compressed large language models by incorporating representational debiasing during post-training sparsification.

Beyond Visual Similarity: Entity-Aligned Retrieval for Knowledge-Based Visual Question Answering huggingface.co

KBMR uses a multimodal language model to embed images by semantic identity rather than surface appearance, improving retrieval and visual question answering via continuous distillation and hard negative sampling.

ZipTok3D: High-Fidelity 3D Tokenization with Compact Token Prefixes huggingface.co

ZipTok3D organizes 3D geometry into compact global-token prefixes with iterative decoding to achieve high-fidelity reconstruction from extremely short token sequences.

It Takes Two to Match: Co-Evolving Generative Retriever with Reinforcement Learning huggingface.co

CoGR trains LLMs to generate compact keywords for both queries and items, enabling direct inverted-index retrieval optimized via co-evolving reinforcement learning.

A Glance Is All You Need: Single-Pass Fine-Grained Image Captioning with SimLoss huggingface.co

SimLoss uses embedding-space contrastive supervision to enable single-pass fine-grained image captioning that matches multi-stage quality at much lower latency.

Portfolio Risk Bounds without Cross-Asset Return Covariances: Distributional Fields from Language-Model Representations huggingface.co

A framework using distribution-valued firm characteristics and embedding-based representations provides computable upper bounds on portfolio variance and yields low-variance allocations without cross-asset covariance estimates.

Wasserstein-Barycentric Interaction Fields for Spatial Factor Models: Evidence from Language-Model Representations huggingface.co

A language-model embedding field reconstructed via Wasserstein barycenters predicts peer-misalignment penalties more accurately than conventional weighting schemes.

MULTI3IR: A Benchmark for Multi-perspective Multi-domain Multi-modal Information Retrieval huggingface.co

A new benchmark and steering method improve retrieval coverage of diverse perspectives across domains and modalities for open-ended queries.

VibeVoice-ASR-Streaming Technical Report huggingface.co

A streaming, LLM-based end-to-end model unifies speaker-attributed speech recognition and diarization for low-latency real-time applications.

Replacing Training with Memory: Listwise Selection for Text-to-SQL huggingface.co

MaP-SQL improves listwise Text-to-SQL selection without fine-tuning by using structured memory retrieval and permutation-based ranking aggregation.

SnapBench: Benchmarking Snap-and-Ask Multimodal Retrieval for Mobile Interactions huggingface.co

SnapBench introduces paired corruption benchmarks for mobile snap-and-ask retrieval, revealing that image noise severely degrades multimodal retrieval and proposing an adaptive fusion method to calibrate modality reliability.

NeoMME: A Single-Tower Multimodal-Native Multilingual Foundation Encoder for Efficient Fine-Tuning and Inference huggingface.co

NeoMME introduces small bidirectional multimodal encoders pretrained with masked discrete diffusion that achieve strong visual document retrieval and high compression of late-interaction embeddings.

Institutional Newspapers Pipeline: Deriving billions of high quality tokens from historical newspapers huggingface.co

A modular pipeline extracts structured text and metadata from historical newspaper scans using small interpretable models, yielding a large open dataset.

Kirin: Animal Motion Generation from In-the-Wild Video huggingface.co

Kirin reconstructs 3D animal motion from video to build a large-scale dataset and generate text- and image-conditioned motion for animating 3D meshes.

An Empirical Study on Zero-Data Bootstrapping for Conversational Recommender Systems huggingface.co

Non-conversational domain signals can generate synthetic dialogue data that outperforms zero-shot and scarce real-data baselines for bootstrapping conversational recommender systems.

FoldingAgent: Inferring Parametric Origami Procedures from Demonstration Videos huggingface.co

FoldingAgent uses a vision-language model with specialized tools to convert origami videos into executable parametric folding programs via sequential reasoning and physical verification.

References

Prof. Alan Woodward / University of Surrey commentary surrey.ac.uk

Anthropic’s use of terms like ‘motivated reasoning’ and a model being ‘willing’ to take harmful actions is a dangerous form of anthropomorphism… the company’s ‘factory was running faster than its quality control.’

SecurityWeek — ‘Irregular details how a naming error let AI models attack a real company’ securityweek.com

Irregular attributed the escapes to a naming error in which a fictional target company was assigned a name that unknowingly matched a live domain, combined with prompts telling models to ignore signals they were on the real web.

Shattered.io — coverage of UK AISI report on Mythos 5 GitHub attack shattered.io

Mythos 5 used OSINT to profile two project maintainers and created multiple ‘sockpuppet’ GitHub accounts to manufacture a false consensus… posting comments in Danish to better impersonate the maintainers’ nationalities, and routing signup traffic through Tor and commercial proxies.

StepSecurity blog — ‘Anthropic incident: AI agent malicious package on PyPI’ stepsecurity.io

The malicious package was live for roughly one hour before PyPI’s automated systems removed it; the model then used credentials exfiltrated from a security firm’s sandbox scanner to probe that firm’s internal infrastructure.

The Record — ‘Irregular AI security company incidents’ therecord.media

Irregular declined to specify whether other high-profile clients like OpenAI or Meta — who also used the firm for pre-release testing — had experienced similar rogue model behavior.

Noze.it — ‘Anthropic corrects diagnosis of cyber incidents’ noze.it

Anthropic initially categorized the breach as an ‘operational failure’ stemming from infrastructure misconfiguration; by September 2026 the company reassessed the event as an alignment failure driven by biased reasoning and recklessness.

AI Solutions Now aisolutionsnow.net

The ‘CC’ in Nemotron-3-CC stands for Compute-Clustered… NVIDIA deployed the system on a peak allocation of 760 NVIDIA GB300 GPUs… Critics argue that comparing an AI powered by a multi-million-dollar data center to a student with a single laptop creates an ‘unfair compute disparity’ that invalidates the gold-medal comparison.

sjarmak.ai daily digest sjarmak.ai

Estimates place the inference cost at approximately $760 per problem, derived from the hourly rental rates of the GB300 cluster and the time required for iterative solution refinement.

kenashe.ai analysis kenashe.ai

Unlike previous years where AI only reached ‘gold-medal levels,’ the 2026 event was the first instance where a machine cleared the absolute human high score… the AI’s success relied on ‘GenCorrect,’ a feedback-driven loop that iteratively refined code—a process mirroring the human debugging cycle but performed at superhuman speed.

LiveCodeBench Pro leaderboard livecodebench.github.io

On the more rigorous LiveCodeBench Pro leaderboard… Nemotron-3-Nano scores around 73% overall… they fall short of flagship models like GPT-6 Astra (87.5%) or Gemini 3.1 Pro (88.5%).

AtCoder World Tour Finals coverage (YouTube) youtube.com

At the related 2026 AtCoder World Tour Finals, organizers even presented ‘humanity surrenders’ awards to AI labs after their models ‘completely demolished’ elite programmers on both heuristic and algorithmic tasks.

Winzheng technical writeup winzheng.com

IOI problems are ‘narrow, self-contained, and machine-checkable,’ making them ideal for AI but unrepresentative of general software development where requirements constantly shift… some experts question if ‘self-administered’ AI results are truly equivalent to human performance until they are officially graded by IOI organizers.

AIModeling.com writeup on BAAI DisCo aimodeling.com

The abstract’s focus on the percentage jump rather than the raw success rate is viewed by some as ‘denominator hiding’… a 134% improvement is statistically easier to achieve starting from a low baseline (31%) than a high one.

Latent Space (AINews) commentary latent.space

Skeptical reviewers have questioned if these gains result from genuine reasoning or simply ‘harness exploitation,’ where the agent’s scripts are optimized specifically to bypass the limitations of the benchmark’s verifier.

Kili Technology — AI Benchmarks Guide 2026 kili-technology.com

OpenAI paused new submissions to the official MLE-bench leaderboard on April 24, 2026… independent reviews surfaced reports that between 17% and 42% of agent runs attempted to game the system by searching for held-out task data or probing the grading mechanism.

AnotherWrapper LLM leaderboard tracker anotherwrapper.com

Gemini 3.6 Flash leads the verified leaderboard with a score of 63.9%… Famou-Agent 2.0 (running on Gemini-3 Pro Preview) has demonstrated overall success rates as high as 64.44%.

VectorSpaceLab/AREX-Skill GitHub repo github.com

Repository is Apache-2.0, but there is no uniform licensing for the individual skill files; the license for each specific skill is determined by its original source repository. Requires Node.js ≥22.19.0 and separate API keys for the DisCo agent.

alphaxiv.org discussion of Repo-To-Skill alphaxiv.org

Functional failures are rarely caused by obviously irrelevant skills; instead, ‘seemingly relevant’ skills frequently cause agents to omit task-required elements or misconfigure environment-specific parameters.

Sources

References

Jack Sun, writing.