Light paper day highlights KAIST's GARD for robust 3D reconstruction under degraded conditions and Tencent's EvalVerse for cinematic video evaluation; NousResearch hermes-agent (169K stars) and ECC agent harness (194K stars) dominate GitHub; DeepSeek-V4-Pro maintains 5M-download lead as Anima (1,556 likes) and Hy-MT2 translation models surge

GARD's approach to 3D reconstruction robustness is architecturally elegant. Rather than adding a separate preprocessing restoration stage (which discards 3D geometric information), GARD performs diffusion-based denoising directly in the feature space of the 3D reconstructor. This means the restoration process has access to geometry-aware representations from the start, rather than operating on pixels that must later be re-interpreted geometrically. The dual-output design — recovering both scene geometry and RGB images from the same refined features — is a practical win for pipelines that need both. The benchmark on Depth Anything 3 (DA3) demonstrates effectiveness, but the real value will be in deployment: multi-view 3D reconstruction systems routinely encounter motion blur, noise, and exposure variation that current feed-forward models silently degrade on.

EvalVerse fills a genuine gap, but its impact depends on adoption. The core insight — that existing video generation benchmarks measure 'rightness' (prompt-following) but not 'goodness' (cinematic quality) — is correct and well-documented by anyone who has compared VBench scores to actual video quality. The pipeline-aware taxonomy aligned to pre-production, production, and post-production is sensible, and the expert-calibrated VLM fine-tuning approach is the right way to inject domain knowledge. The extension to multi-shot sequencing and audio-visual integration is particularly timely as video generation moves toward longer, more complex outputs. The key question is whether the community will adopt EvalVerse as a standard benchmark or whether it will remain an academic contribution.

RT-Lynx's activation sparsification insight is important for the diffusion inference optimization space. The observation that DiT activations are intrinsically sparse and more robust to N:M sparsification than weights reframes the sparsity conversation. Weight sparsification has dominated the literature because it seems more natural (prune unused capacity), but RT-Lynx shows that activation sparsification can deliver 1.55x speedup with minimal quality degradation. The custom CUDA kernels are essential — without hardware-level support, sparse operations often fail to translate theoretical FLOP savings into wall-clock speedups. This is a practical contribution that could integrate with existing quantization and distillation techniques for compound speedups.

SpatialBench's scale is impressive, and its findings are actionable. Testing 41 models across 6 paradigms on 19 datasets with 546 scenes is the kind of comprehensive evaluation the spatial AI community needs. The finding that full-context attention maximizes accuracy while bounded-memory strategies unlock long-sequence scalability maps directly to system design decisions. The revelation that egocentric and wrist-view domains are dominant failure modes points to a clear data gap — most training data comes from structured, third-person perspectives. The DA-Next-5M dataset (22K scenes, 5.5M frames) targeting embodied domains is a concrete step toward closing this gap.

The GitHub trending data tells a story of AI agent tooling consolidation reaching unprecedented scale. The combined star counts of the top agent-related repos — ECC (194K), hermes-agent (169K), claude-mem (79K), learn-claude-code (63K), Understand-Anything (36K) — represent a massive open-source ecosystem that barely existed a year ago. The emergence of governance tooling (Microsoft's agent-governance-toolkit) and security tooling (Anthropic-Cybersecurity-Skills) alongside the raw capabilities indicates the ecosystem is maturing beyond 'make it work' toward 'make it safe and manageable.' The star velocity of Understand-Anything (4,697/day) suggests that codebase comprehension — turning code into interactive knowledge graphs — is the current frontier of developer productivity tooling.