absorb.md

Kai

Chronological feed of everything captured from Kai.

paper / kaicathyc / Jun 11

InternVideo3: Enhancing Long-Horizon Multimodal Understanding with MCR and Efficient Context Handling

InternVideo3 addresses the underexploration of long-horizon multimodal tasks in open-source foundation models by introducing Multimodal Contextual Reasoning (MCR). This framework employs a closed-loop process over an evolving context for evidence accumulation and verification. To enhance efficiency, M^2LA compresses KV-cache states while preserving token streams. The model's effectiveness is demonstrated through a staged training process and strong performance on benchmarks, with further instantiation as a video agent showcasing evidence-grounded behavior.

multimodal-reasoningfoundation-modelsvideo-understandingagentic-aicontext-handlingcomputer-vision
“InternVideo3 enhances long-horizon multimodal tasks by introducing Multimodal Contextual Reasoning (MCR), which frames video understanding as evidence accumulation and verification within a closed-loop context”
paper / kaicathyc / May 22

LLM-Generated Test Suites Are Too Weak to Catch Subtle Bugs: Evidence from SWE-Mutation

SWE-Mutation is a benchmark designed to evaluate the discriminative quality of LLM-generated test suites by introducing systematically mutated (near-correct) solutions that attempt to pass validation undetected. The core finding is that current LLMs — including frontier models like DeepSeek-V3.1 — produce superficial test suites with critically low verification and detection rates, exposing a fundamental bottleneck for LLM-based software engineering and reinforcement learning feedback pipelines. The authors also demonstrate that an agentic mutation strategy generates significantly more realistic and harder-to-detect mutants than conventional mutation testing methods.

llm-evalssoftware-engineeringtest-generationmutation-testingbenchmarkingcode-llmsreinforcement-learning
“Even the best-performing LLM (DeepSeek-V3.1) achieves only 10.20% verification and 36.15% detection rates on SWE-Mutation, indicating that current LLMs cannot generate reliably discriminative test suites.”
paper / kaicathyc / May 22

Ontology-Grounded Knowledge Graphs as the Authoritative Runtime Backbone for Circular Manufacturing Systems

Conventional manufacturing IT architectures, built for stable, deterministic processes, fail to handle the variability and uncertainty inherent in circular manufacturing, where reintroduced products have heterogeneous and unknown states. KAPPS addresses this by centering a Cyber-Physical Production System (CPPS) on an ontology-grounded knowledge graph that serves not merely as an integration layer but as the factory's authoritative write-time state. The architecture adds a semantic interface layer, constraint enforcement modules, and event-driven planning to enable dynamic process reconfiguration, runtime reasoning, and human-machine knowledge exchange — validated through anomaly detection and modular conveyor constraint enforcement use cases.

knowledge-graphscyber-physical-systemscircular-manufacturingontologysmart-factoryindustrial-aisemantic-reasoning
“Conventional manufacturing IT architectures cannot adequately represent or manage the uniqueness of individual components at runtime, making them unfit for circular manufacturing.”
paper / kaicathyc / May 22

Priority Ranking: A Low-Cost Direct Evaluation Method for LLM-Based Harness Optimizers

Current harness optimization research evaluates optimizer agents solely through end-task performance gains of target agents, obscuring whether improvements stem from informed decisions or random trial-and-error. This paper introduces "priority ranking" — a step-level evaluation design where optimizers rank harness components (e.g., tools) by their potential impact, enabling direct assessment without expensive rollouts. The authors demonstrate that ranking performance correlates with actual multi-step optimization ability, making it a reliable proxy metric. The approach is grounded in Shor, a curated benchmark of 182 human-verified optimization scenarios across diverse domains and design stages.

ai-evalsagent-optimizationllm-agentsbenchmarkautomated-agentsharness-optimization
“Existing harness optimizer evaluations are indirect and fail to capture optimizer behavior at intermediate steps, leaving it unclear whether gains are from informed decisions or trial-and-error.”
paper / kaicathyc / May 22

Optimal Reed-Muller Code Testing Achieved Under Adversarial Online Erasures

Kelman, Meir, and Zheng introduce *semi-sample-based testers* — a hybrid between fully adaptive and purely random-query testers — to address Reed-Muller code property testing under the online-erasure adversary model, where an adversary can erase up to *t* input points after each query. They prove these testers achieve optimal query complexity in this adversarial setting, improving on Minzer and Zheng (SODA 2024). As a secondary result, semi-sample-based testers are shown to generalize to lifted affine-invariant codes (Guo, Kopparty, Sudan 2013), yielding the first known testers for those codes under online erasures.

reed-muller-codesproperty-testingonline-adversaryerror-correcting-codesquery-complexityprobabilistically-checkable-proofsalgorithms-theory
“Semi-sample-based testers achieve optimal query complexity for testing Reed-Muller codes in the online-erasure model.”