Proof, Then Trust: Lean Certifies an AI-Generated QAOA Result
A machine-checked QAOA proof shows how generative systems can search broadly while deterministic verification controls what gets accepted.
Cognaptus Insights
Industry thoughts, trends, and commentary.
Rank the Work, Not the Model: Meta-Benchmarks for Bank LLM Screening
A financial-services meta-benchmark turns public LLM results into domain-specific screening evidence—without pretending that a ranking is deployment approval.
Refusal Is Not a Result: Vera Tests What Agents Actually Changed
Vera reframes agent safety evaluation as reproducible software testing built around observable effects, adaptive attacks, and deterministic verification.
Name the Speaker, Then Ask the Plot: Selective Reasoning for Drama Transcripts
DramaSR-532K shows that multimodal reasoning is most valuable when acoustic speaker attribution becomes uncertain, not as a replacement for the acoustic pipeline.
Refactor, Then Run It: SemaDiff Tests Whether Behavior Actually Changed
SemaDiff shows how generated cross-version callers can turn refactoring review from a structural guess into high-precision behavioral evidence.
The Mask Is Not the Model: MMIR-TCM Makes Clinical Memory Inspectable
MMIR-TCM shows how staged perception, reporting, and retrieval can support auditable TCM assistance—without establishing autonomous prescribing.
No Runtime, No Signal: Java Energy Prediction Beyond Static Metrics
Static code metrics barely predict Java method energy use; lightweight execution timing offers a better signal, but not a production-grade energy meter.
Search the Graph, Not the Model: RSF-GLLM Separates Traversal from Generation
RSF-GLLM shows how a dedicated graph reasoner can traverse weakly related bridge entities, expose auditable evidence paths, and reduce reliance on repeated LLM calls.
The Missing Present Is a Distribution: DUPO for Delayed Control
DUPO shows why delayed controllers should evaluate actions across several plausible current states instead of trusting one reconstructed present.
Merge Before You Stop: Checkpoint Choice Depends on the Operator
A controlled model-merging study shows why expert checkpoints should be selected jointly with the merge operator rather than fixed by standalone validation loss.