The Generalization Stack: Why HAR Robustness Is a Pipeline Property
A 410,400-experiment HAR benchmark shows why deployment robustness depends on the full training-and-selection pipeline, not a single domain-generalization method.
Cognaptus Insights
Industry thoughts, trends, and commentary.
The Guard Is Already in the Draft: Reusing Speculative Decoding for LLM Monitoring
Speculative Probing shows how systems already paying for speculative decoding may reuse that infrastructure for sequence-aware monitoring at much lower incremental cost.
The KL You Weren’t Watching: Moving RLVR Stability to the Query Side
ERPO shifts explicit drift control from model responses to the queries themselves, aiming to preserve exploration while improving RLVR stability.
Turn Down the Work, Not Just the Clock: Governing On-Device LLM Power Across Hardware and Decoding
PELM shows why efficient on-device LLM inference is a joint control problem spanning processor frequency, model computation, latency, thermal headroom, and output quality.
Watching a Signal the Model Can Move
Activation probes are not automatically tamper-resistant: a new benchmark shows that current LLMs can deliberately reshape some of the internal signals those monitors read.
Keep the Rollout Honest: Miles Treats Throughput and Fidelity as One System
Miles v0.1 shows why frontier post-training teams should optimize utilization, data freshness, numerical fidelity, memory placement, and weight movement as one coupled system.
Left of Whom? Spatial Agents Need More Than an Explicit Viewpoint
POVBench shows that explicitly supplying a human viewpoint does not remove the hardest parts of observer-relative spatial reasoning, changing how embodied AI teams should diagnose and remediate failures.
Local Evidence, Global Rule: When Knowledge Graph Embeddings Generalize Too Far
PogRE shows how knowledge-graph models can turn sparsely supported relational patterns into overly broad rules—and how generalization scope can instead grow with accumulated evidence.
Safety Without the Data Lake: Federating the Guard, Not the Traces
FGLGuard shows why multi-agent safety may require local adaptation even when organizations cannot centralize the sensitive traces needed to train it.
Same Count, Different Bugs: Build LLM Security Scans as an Evidence Funnel
Bugstone-E2E shows how CVE patch history can become reusable detection knowledge—and why LLM security findings need staged verification rather than one-shot trust.