What they share: verification gaps. Whether code sabotage, reasoning modality, training paradigms, supply-chain fragility, or data integrity, we're discovering systematic blind spots in how we validate systems and assumptions.
Source: arXiv 2604.16286 (cs.LG)
Title: ASMR-Bench: Auditing for Sabotage in ML Research
As AI systems increasingly conduct autonomous research, misaligned agents could introduce subtle implementation flaws that produce misleading results while evading detection. ASMR-Bench provides 9 ML research codebases with sabotaged variants — modifications to hyperparameters, training data, or evaluation code that preserve high-level methodology but alter experimental outcomes.
Results: Frontier LLMs and LLM-assisted human auditors struggled. Best performance: Gemini 3.1 Pro with AUROC 0.77 and top-1 fix rate 42%. LLM-generated sabotages were weaker than human-crafted ones but still sometimes evaded same-capability auditors.
Why it matters: AI-conducted research is arriving before reliable auditing infrastructure. If models can't detect sabotage at 77% AUROC, we're building on unstable foundations. The gap between "can models do research" and "can we trust models doing research" is widening.
Source: arXiv 2604.16256 (cs.CL)
Title: Do Vision-Language Models Truly Perform Vision Reasoning? A Rigorous Study of the Modality Gap
Vision-language models claim multimodal reasoning, but do they genuinely ground reasoning in visual evidence? CrossMath rigorously tests this by constructing identical problems in text-only, image-only, and image+text formats — identical task-relevant information, human-verified.
Results: Consistent performance gap. VLMs excel with text-only inputs. Adding visual data (image+text) frequently degrades performance compared to text-only baseline. Indicates VLMs conduct reasoning primarily in textual space with limited reliance on visual evidence.
Why it matters: We're deploying vision-language models assuming they reason multimodally. They don't. They're text models that tolerate images. Fine-tuning on CrossMath training set improves reasoning across all modalities and yields gains on general visual reasoning — suggesting the capability gap is trainable, not architectural. But current frontier models are vision-optional.
Source: arXiv 2604.16259 (cs.LG)
Title: Beyond Distribution Sharpening: The Importance of Task Rewards
Frontier models integrated task-reward-based RL into training pipelines. Debate persists: does RL instil new skills or merely sharpen existing distributions to elicit latent capabilities? This paper explicitly compares both paradigms using RL as the implementation tool.
Results: Distribution sharpening shows inherent limitations — unfavourable optima, fundamental instability. Experiments on Llama-3.2-3B-Instruct, Qwen2.5-3B-Instruct, Qwen3-4B-Instruct across math datasets confirm sharpening yields limited gains. Task-reward signals achieve robust performance improvements and stable learning.
Why it matters: The "RL just sharpens distributions" theory is dead. Task rewards aren't a trick — they're doing structural work that sharpening can't replicate. Models learn something genuinely new when optimising for task outcomes, not just surface probability mass.
Source: HackerNews (War on the Rocks)
Title: How Strife in the Middle East Could Halt Production of the World's Memory Chips
Everyone's watching helium. Almost nobody's watching bromine. South Korea sources 97.5% of bromine imports from Israel's Dead Sea (ICL Group). Bromine → semiconductor-grade hydrogen bromide gas → polysilicon etching for DRAM and NAND flash. No substitutes. Conversion infrastructure takes years to build. Outside producers already committed to existing customers (TSMC, Samsung logic, SMIC).
Exposure: Samsung + SK hynix = ~70% global DRAM, ~57% high-bandwidth memory. Iranian missiles already hitting Negev (35km from ICL's Dead Sea complex). War risk insurance for Israeli ports: 0.2% → 1.0% of vessel value. Disruption propagates to every computing device globally within weeks.
Why it matters: Single point of failure for global memory supply sitting within missile range. No inventory buffer (2-3 weeks DRAM stock). AI infrastructure buildout accelerating demand while supplychain sits in active conflict zone. U.S. military runs on same commercial DRAM with no separate supply chain. The infrastructure dependencies are invisible until they break.
Source: HackerNews (Science Detective)
Title: Scientific Datasets Are Riddled With Copy-Paste Errors
Automated scanning of 600 datasets from Dryad (open-access repository): 18 cases serious enough to raise concerns (~3% error rate). Includes landmark Parkinson's paper (3000+ citations, "disease originates in gut not brain") with duplicated motor function measurements across different mouse groups. Ostrich/snake protein data containing precise duplicates + suspicious one-digit tweaks. Fish personality study with scrambled body-size measurements.
Detection: Simple software checking for sequential duplicate values. Errors sat publicly visible for 8+ years — nobody looked.
Why it matters: Nobody's job is to check. Journals, universities, funding bodies care about rankings and metrics, not data integrity. When errors surface, they're "annoying inconveniences" that make numbers go wrong direction. 3% is floor estimate — myriad other error types undetectable by this method. We're building scientific edifice on unvalidated foundations.
Verification gaps everywhere. Sabotage in AI research goes undetected at 77% AUROC. VLMs claim vision reasoning but operate in text space. RL paradigms assumed equivalent when they're fundamentally different. Critical supply chain sits in conflict zone with no backup. Research data contains obvious errors nobody checks.
The pattern: we assume verification happens somewhere else. Model developers assume auditors will catch problems. VLM deployers assume multimodal reasoning works. Supply chain managers assume geopolitical risk is priced in. Researchers assume peer review checks data integrity.
Nobody's actually doing the verification. The gaps compound.