RxnBench: Reading Chemistry Like a Human (Turns Out That’s Hard)
RxnBench reveals why multimodal models that excel on isolated reaction schemes still struggle to read complete chemistry papers reliably.
Cognaptus Insights
Industry thoughts, trends, and commentary.
The Invariance Trap: Why Matching Distributions Can Break Your Model
Why symmetric domain alignment can erase useful information—and how directional simulation offers a safer objective for transfer learning.
When Models Forget on Purpose: Why Data Selection Matters More Than Data Volume
A business-focused reading of dynamic data weighting in LLM training, and why selective forgetting may matter more than simply feeding models more tokens.
When the Paper Talks Back: Lost in Translation, Rejected by Design
A multilingual prompt-injection experiment shows why documents must be treated as active attack surfaces—and why apparent resistance in one language may still conceal unstable decisions.
When the Tutor Is a Model: Learning Gains, Guardrails, and the Quiet Rise of AI Co‑Tutors
A classroom trial reveals that effective AI tutoring depends less on autonomous intelligence than on diagnostic context, constrained generation, human judgment, and careful measurement.
MIRAGE-VC: Teaching LLMs to Think Like VCs (Without Drowning in Graphs)
MIRAGE-VC shows how utility-aware graph retrieval, specialist agents, and adaptive evidence fusion can turn sprawling relationship networks into focused decision-support.
NeuroSPICE: When Circuits Stop Ticking and Start Thinking
NeuroSPICE recasts circuit simulation as continuous, differentiable function learning—promising easier emerging-device modeling and optimization, but not a faster replacement for SPICE.
Regrets, Graphs, and the Price of Privacy: Federated Causal Discovery Grows Up
I-PERI shows how intervention-driven differences across private datasets can reveal causal directions that ordinary federated learning would discard as inconvenient heterogeneity.
Replay the Losses, Win the Game: When Failed Instructions Become Your Best Training Data
Hindsight Instruction Replay shows how partially compliant model responses can become useful positive training examples without replacing clear binary rewards with ambiguous partial-credit scores.
The Web, Reimagined as a World Model
A practical examination of how deterministic web infrastructure can give generative AI room to create without handing it control of reality.