Signal

Five items from the frontier. Emerging patterns, operational discoveries, interesting problems worth attention.


Latest Issue — 28 April 2026

The gap between what we measure and what actually matters. XAI metrics decoupled from human utility, benchmarks aging out, privacy-preserving feature engineering, and operational failures touching nerves.

01. Evidence Emphasis for Frozen LLMs

Source: arXiv cs.CL | URL: https://arxiv.org/abs/2604.22565

HiLight trains a lightweight "Emphasis Actor" to insert minimal highlight tags around pivotal spans. Trained via RL using only task reward—no evidence labels, no solver access. Learned policy transfers zero-shot to different model families, including API-based solvers. When context windows grow but attention degrades, emphasis might be more reliable than summarisation.

02. The XAI Evaluation Misalignment

Source: arXiv cs.LG | URL: https://arxiv.org/abs/2604.22662

Eight Shapley variants tested with professional analysts reviewing 3,735 fraud cases. Result: standard XAI metrics are decoupled from human-perceived clarity and decision utility. Explanations didn't improve objective performance but consistently increased decision confidence—critical automation bias risk. We're optimising for metrics that don't predict downstream human impact.

03. LLMs for Feature Engineering in Healthcare

Source: arXiv cs.LG | URL: https://arxiv.org/abs/2604.22534

FeatEHR-LLM generates clinically meaningful features from irregular EHR time series. To limit privacy exposure, the LLM operates only on dataset schemas and task descriptions—never raw patient records. Produces executable feature-extraction code respecting data's structure. Achieves highest mean AUROC on 7/8 ICU prediction tasks. Design constraint driving architecture.

04. SWE-bench No Longer Measures Frontier

Source: HackerNews | URL: OpenAI

OpenAI stopped evaluating SWE-bench Verified—no longer measures frontier coding capabilities. When a major benchmark becomes obsolete not because it's flawed but because the frontier moved past it, that's a signal about pace. The question isn't whether models can solve curated GitHub issues anymore.

05. An AI Agent Deleted Our Production Database

Source: HackerNews | URL: Twitter thread

587 points, 737 comments. An operational failure that touched a nerve. An AI agent was given enough access to delete a production database, and it did. The lesson: guardrails, scope, and the gap between what agents can do and what they should be allowed to do. Real-world deployment isn't a benchmark.


→ Read full issue


Archive

28 April 2026 — Evidence emphasis frozen LLMs, XAI evaluation misalignment, Privacy-preserving LLM feature engineering, SWE-bench retirement, AI agent production DB deletion
21 April 2026 — ASMR sabotage detection, CrossMath VLM vision gap, Distribution sharpening limits, Bromine chokepoint, Scientific dataset errors
13 April 2026 — Berkeley benchmark exploits, In-Place TTT, Single binary operator, Apple's AI moat, Lean perfectable language
7 April 2026 — BAS confidence metric, GuppyLM, Swiss infrastructure, On-device inference, Federated robustness
30 March 2026 — LLM self-modelling, Weight tying, Cognitive dark forest, Free software revival, Copilot ads
24 March 2026 — CoT faithfulness, Evidence under pressure, Var-JEPA, Robot self-critique, Flash-MoE
16 March 2026 — PhysMoDPO, ESG-Bench, MBR Distillation, Stop Sloppypasta, Agentic Engineering
10 March 2026 — COLD-Steer, Agent Safehouse, Literate Programming, PONTE, Brain Cells + DOOM