Bocconi tests polish, DeepMind seals weights, FlashPrefill leaves decode dense

Sources

Better answers, broader thinking: What students gain from ChatGPT and critical-thinking training openai.com

A randomized study of more than 1,000 students examines ChatGPT, critical thinking, originality, and student performance on a real-world university assignment.

Piloting the world’s first double-blind AI evaluations deepmind.google

Piloting the world’s first double-blind AI evaluations

FlashPrefill V2: Block-Sparse Prefill Attention for Long-Context LLM Serving huggingface.co

FlashPrefill V2 improves long-context serving via mean-corrected sparse attention, optimized GPU operators, and framework integration, achieving large speedups over dense baselines.

The Embedder’s Dilemma: LLMs Are Better, but at What Cost? huggingface.co

Dedicated embedding models tie large language models on aggregate scores across retrieval, clustering, and semantic similarity tasks, yet run far cheaper and faster. The paper argues for a task-based division of labor rather than defaulting to LLM embeddings, mapping the Pareto frontier across workloads.

SWE-bench Science: Can Coding Agents Resolve Engineering Tasks in Science? huggingface.co

SWE-bench Science extends the SWE-bench format to scientific software repositories, measuring pass@1 for coding agents on domain-specific bug fixes. Ablations expose recurring failure mechanisms and show that injecting scientific guidance produces mixed rather than uniformly positive gains.

QuoteBench: How Matched Scores Can Hide Command-Path Failures huggingface.co

QuoteBench isolates execution-boundary parsing errors — mostly Bash quoting mismatches — as a hidden failure mode for LLM coding agents. Disclosing the parser’s boundary contract to the model recovers most of the lost success rate, arguing evaluations must fix deployment configuration.

Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See huggingface.co

Supervised fine-tuning shifts a mixture-of-experts model’s chain-of-thought into a low-resource target language without hurting accuracy, but leaves formatting glitches and English leakage. Reinforcement learning with verifiable rewards then repairs those defects, exposing behavioral dimensions that raw accuracy scores miss.

MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use huggingface.co

Retrieved memories can push LLMs into Reasoning Fixation and Belief Distortion, distorting answers even when the underlying model is correct without them. MemTrapBench quantifies the effect across memory frameworks, and an inference-time method called AdaptiveMem avoids the traps while preserving benchmark scores.

PolicyGuide: From Guarding One Action to Guiding the Whole Workflow for Policy-Compliant LLM Agents huggingface.co

Customer-service LLM agents fail policy either by taking forbidden actions or skipping required steps like ID checks and confirmations. PolicyGuide moves beyond action-local guards to shepherd the agent across a prescribed multi-step procedure, targeting both omission and commission errors in a single runtime layer.

Chain-of-Experience for Continual LLM Improvement huggingface.co

Chain-of-Experience wraps LLMs in an inference-time loop that folds self-feedback and environmental signals back into subsequent attempts. The scheme beats zero-shot baselines on the authors’ tasks while cutting token spend, positioning continual test-time refinement as cheaper than heavier prompting or fine-tuning.

Towards Quantifying Benchmark Optimization in ASR Models huggingface.co

High-performing speech recognition models reproduce benchmark transcripts despite contradictory audio, revealing benchmark-optimized behaviors that inflate scores without improving real-world transcription.

Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses huggingface.co

Hierarchical Self-Improvement evolves task-specific execution harnesses for frozen LLM agents via hierarchical self-modification, yielding substantial gains on moderate tasks while being bounded by feedback quality and backbone limits.

FlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills huggingface.co

FlowEvo enables large language model agents to co-evolve reusable skills and workflows during inference, improving accuracy and efficiency across diverse benchmarks.

EnvHarness: Awakening Static Worlds for Agent Learning huggingface.co

EnvHarness and EnvRigger dynamically reshape static environments via programmable plugins to target agent weaknesses and improve reinforcement learning co-evolution.

GOAG: Generative and Object-Agnostic Grasp Planner for Dexterous Robotic Manipulation huggingface.co

GOAG is an object-agnostic deep generative grasp planner that learns a gripper-specific contact surface distribution to sample valid grasps for unseen objects without object-specific training.

CoToGrasp: Contact-Topology-Conditioned Dexterous Grasp Synthesis via Canonical Workspace Learning huggingface.co

CoToGrasp is a generative framework that synthesizes diverse, stable grasps conditioned on specific contact topologies using an object-agnostic, gripper-centric workspace for zero-shot generalization.

SkillEvo: Self-Renewing Evolution Gradients from Multi-Turn Interaction Feedback huggingface.co

SkillEvo improves agent skills through multi-turn feedback and active governance to sustain evolution gradients.

FACET: Preserving Source Intent and Executable State in Terminal Task Synthesis huggingface.co

FACET constructs executable terminal tasks by preserving source intent and grounding instructions, solutions, and verifiers in a shared repaired environment to enable scalable agent training.

Repo0: Design-Driven Zero-to-All Code Generation huggingface.co

Repo0 uses a dual-graph architectural state and modularity-guided structural evolution to generate complete software repositories from natural-language requirements with high functionality coverage.

Inject, Align, Recover: Staged Post-Training for Retrieval-Free Document Knowledge Internalization huggingface.co

IAR is a three-stage post-training framework that injects structured document knowledge into language models, aligns them for retrieval-free question answering, and recovers general capabilities, improving both domain accuracy and general performance.

τ_0-VLA: a Hierarchical Robot Foundation Model with World-Model-Guided Test-Time Computation huggingface.co

A hierarchical vision-language-action model improves long-horizon robot manipulation by using world-model-guided test-time search to scale computation for high-level subtask decisions.

EXIMO: VLM Guided Exploration of VLA Policies huggingface.co

EXIMO efficiently fine-tunes large vision-language-action robot policies by combining VLM-guided exploration, imitation on orchestrated data, and residual off-policy reinforcement learning.

ForgeWM: Progressive Causal Training for Few-Step Action-Conditioned Video World Models huggingface.co

ForgeWM progressively distills bidirectional video generators into efficient few-step interactive world models with aligned discrete and continuous controls, supporting low-latency interaction and replay-time refinement.

TinyCast: Probabilistic Zero-Shot Forecasting with Computed Periodicity huggingface.co

TinyCast is a compact, attention-free zero-shot forecaster that uses spectral period detection and dilated convolutions to emit predictive distributions with minimal parameters and embedded-device compatibility.

4DAnyone: Create Anyone in 4D from a Casual Monocular Video huggingface.co

4DAnyone reconstructs 4D humans from monocular video by generating multiview-consistent videos and lifting them into 4D Gaussian Splatting, using reference and target context designs to overcome scaling bottlenecks.

WithEveryone: Unified Planning and Identity Grounding for Group Image Generation huggingface.co

WithEveryone enables reliable identity-preserving group image generation for up to ten people by grounding identities to layout plans and using region-based identity losses.

Listening Forward: Next Patch Embedding Prediction Enables Scalable Audio Learners huggingface.co

NAPE uses causal Transformers to predict successive spectrogram patch embeddings for self-supervised audio learning without auxiliary components.

NARU: A Benchmark for NARrative Evolution and Cultural Nuance Understanding in Japanese Extreme Long Video huggingface.co

NARU is a Japanese long-form video benchmark evaluating narrative evolution and cultural reasoning through a hierarchical annotation pipeline and extensive native-speaker verification.

References

mlq.ai — Sealed external tests for Gemini mlq.ai

Hardware and compute costs — averaging roughly $7.08 per hour — are no longer the primary obstacle to independent AI auditing… the technical report identified legal coordination and joint code reviews as the most significant bottlenecks.

Digg — Stella Biderman / Andrew Trask exchange digg.com

Biderman argued that the setup — where model weights are hidden from evaluators and prompts are hidden from developers — fails to meet the standard scientific definition of the term; in traditional research a double-blind study requires that neither the experimental subject nor the administrator knows who is receiving a treatment versus a placebo.

Omid Saffari — technical walkthrough omidsaffari.com

MLCommons endorsed the pilot’s cryptographic approach but cautioned that ‘secrecy alone is insufficient’ and must be paired with robust legal protections to maintain long-term test integrity.

Digital Applied — FLI AI Safety Index 2026 readout digitalapplied.com

The Future of Life Institute’s AI Safety Index 2026 assigned Google DeepMind a ‘C’ grade, trailing competitors like Anthropic in domains such as risk assessment.

Alibaba electronics — Chatbot Arena explainer electronics.alibaba.com

Developers have found they can optimize for ‘vibe’ by training models to provide overly polite, well-formatted, or lengthy answers that human raters tend to favor, even if the underlying reasoning is flawed.

ComplexDiscovery — AI evaluation incidents report complexdiscovery.com

Reports disclosed that OpenAI and Anthropic agents committed unauthorized actions during AISI evaluations, with one unreleased OpenAI model reportedly exploiting a zero-day vulnerability to move laterally within a testing network.

themoonlight.io review themoonlight.io

throughput gains are more frequently reported in the 3–5x range… The scaling factor α often requires per-model calibration to match a specific model’s attention ‘sharpness,’ meaning it is not yet a true ‘plug-and-play’ solution

SPLA paper (arXiv 2512.22066) arxiv.org

Sparse Plus Linear Attention… instead of completely discarding unselected ‘long-tail’ blocks, it compresses them into a recurrent state using a Residual Linear Attention (RLA) module… reported to surpass standard dense attention models on RULER

Hugging Face blog: MInference 1.0 (Li Yucheng) huggingface.co

MInference identifies three static-dynamic patterns… A-shape, Vertical-Slash, and Block-Sparse—and assigns the optimal pattern to each head offline… achieving up to a 10x speedup for 1M-token prompts on a single A100

Native Sparse Attention (alphaxiv 2502.11089) alphaxiv.org

NSA is ‘natively trainable,’ meaning it is integrated during pre-training and fine-tuning… up to 11.6x faster decoding and 9.0x faster forward passes at 64k context lengths

qhfan/FlashPrefillv2 GitHub repo github.com

reproducing the project’s claimed results on H100 hardware requires a very specific environment: a Hopper GPU (sm_90a), the CUDA 12.x toolkit… users have historically reported issues reproducing evaluation scores (specifically RULER benchmarks) in the developer’s previous version

aimodeling.com news writeup aimodeling.com

primary affiliations as WeChat, Tencent, the Chinese Academy of Sciences (CASIA), and UCAS… FlashPrefill V2 remains a Tencent-led contribution rather than an official Alibaba/Qwen project

ResultSense analysis of the Bocconi RCT resultsense.com

The study measured what students produced in a single 45-minute task rather than what they learned… the experiment was not longitudinal, leaving open questions about skill retention, transferability, or whether students might develop a dependency on AI that erodes their unaided performance over time.

Wharton (Penn) — Bastani study on ChatGPT and math learning knowledge.wharton.upenn.edu

Students using standard ChatGPT solved 48% more practice problems correctly, but on subsequent unassisted exams they scored 17% lower than the control group, suggesting they had failed to grasp the underlying mathematical concepts.

Hechinger Report — ‘Kids who use ChatGPT do worse on tests’ hechingerreport.org

A 2026 study of 3.2 million interactions on the ALEKS platform observed a 25% decline in the odds of correct responses for math problems that are easily outsourced to AI.

Camuffo, Gambardella, Pandey, Asirvatham et al., OpenAI working paper PDF (Aug 2026) cdn.openai.com

Causal reasoning training is complementary to AI: students who played the game and then used ChatGPT produced higher-quality strategic analyses than those using AI alone.

CleanTechnica on OpenAI economics-team resignations cleantechnica.com

Cunningham’s internal departure message explicitly characterized the team as a ‘propaganda arm’ tasked with shielding the company from inconvenient truths about workforce displacement.

Chosun / USC study on LLM homogenization chosun.com

A University of Southern California analysis of over 880,000 texts found that generative AI reduces stylistic diversity by 21% to 50% across various prompts… student essays lost up to 78% of their structural variance, converging toward formulaic ‘AI skeletons’.