2026-09-11

TL;DR - COBRA-Skills optimizes reusable LLM-agent skills by combining contextual-bandit prioritization with execution-feedback-driven skill evolution. It reports leading average performance across multiple benchmarks while cutting optimization costs and data requirements.

• Treats skill optimization as budgeted sequential optimization over a dynamically evolving candidate pool.
• Selectively evaluates promising or informative skills rather than exhaustively testing candidates.
• Across six agent benchmarks and three target models, it achieves the strongest average performance among compared methods.
• Reduces optimization cost by 55–58% versus SkillOpt while using 50 unique optimization examples per benchmark.

TL;DR - Mr.LHDR is a multimodal benchmark for testing deep-research agents on long, dependency-heavy evidence chains. Its results show that current systems struggle to maintain consistent reasoning across many interdependent steps, and final-answer accuracy can overstate true research success.

• Tasks require an average of 12.1 intermediate conclusions and a mean dependency depth of 10.4.
• The benchmark includes consequential non-text evidence such as images, maps, PDFs, charts, tables, logos, and video frames.
• The strongest evaluated system reached 43.1% Overall Accuracy but only 34.3% Strict Accuracy.
• Removing images lowered the Dependency-Aware Checklist Score by 12.6 points, while strict accuracy declined as reasoning chains lengthened.

TL;DR - Sci-MMR is a 235-task benchmark for evaluating whether multimodal research agents can perform multi-step scientific reasoning with complete, traceable evidence. Results show that answer accuracy overstates evidence-grounded capability by more than 20 percentage points.

• Tasks span four scientific disciplines and use argument graphs linking claims, citations, visual evidence, and supporting figure regions.
• Evidence acquisition caused 57.2% of failures; cropping improved performance by 4.5 points, while gold evidence raised accuracy by up to 37.0 points.
• Evidence integration caused 31.8% of failures, showing that models also struggle to derive correct conclusions from available evidence.
• Even with gold evidence, the strongest evaluated model reached only 69.1% accuracy on the hardest tasks.

TL;DR - AssayLoop combines a transformer-based acquisition policy with LLM-derived biological priors to prioritize CRISPR perturbations across sequential experiments. It discovers 27.7% of hits while testing about 5% of candidates, potentially making constrained biological screens substantially more efficient.

• AssayBench-Loop includes 1,389 CRISPR screens spanning five phenotype categories.
• AssayFormer learns adaptive acquisition strategies from historical screens and updates them using experimental feedback.
• An adaptive handoff integrates LLM priors, while AssayLLM extends the approach through task-specific LLM post-training.
• On temporally held-out screens, AssayLoop achieved 5.67× enrichment over random selection and transferred to phenotype categories excluded from training.

TL;DR - Negative Self-Distillation (NSD) improves LLM reasoning by teaching models to diverge from self-generated flawed reasoning rather than imitate privileged, artificially confident solutions. This preserves exploratory and self-corrective behavior that conventional on-policy self-distillation may suppress.

• NSD requires neither ground-truth answers nor external supervision; the model generates a question-specific negative persona, such as a “careless reasoner.”
• A dynamic gating mechanism isolates reasoning-critical tokens so training targets behavioral flaws without degrading foundational language abilities.
• The approach addresses the confounding of flawed reasoning with ordinary linguistic tokens that makes naive unlearning objectives risky.
• NSD reportedly outperforms on-policy self-distillation and other label-free, self-bootstrapping reinforcement-learning baselines.

TL;DR - Recursive Code World Models reconstruct complex, executable 3D scenes from a single image using compositional scene programs and a recursively invoked vision-language coding agent. The approach improves fine-detail reconstruction while preserving scene-wide geometry and spatial relationships.

• Uses a global-local-global process: establish the full scene, recursively refine unresolved parts, then revisit overall composition.
• Shares a reference-aligned camera projection across recursion levels for consistent geometry.
• Parent-level revisitation corrects boundaries, spatial relationships, and shared errors introduced during local refinement.
• Outperforms prior code-based image-to-scene reconstruction methods; ablations indicate deeper recursion can improve fine-scale reconstruction.

TL;DR - World in World is a training-free control interface for frozen autoregressive video world models that enables viewpoint-controlled rerendering, long-horizon revisiting, and human-motion transfer. It matters because it unifies several control requirements without task-specific modules or additional model training.

• Converts video observations, scene projections, geometry renderings, and retrieved generated states into camera- and time-labelled visual evidence.
• Uses correspondence routing based on persistent point identities and geometry to align target-view queries with source-video tokens.
• Introduces evidence-wise attention classifier-free guidance to regulate each auxiliary evidence channel independently within one denoising pass.
• Evaluates perceptual quality, temporal consistency, and camera-following accuracy across diverse viewpoint changes.

TL;DR - SenseNova-U1.5 is an 8B mixture-of-transformers model that unifies visual understanding, reasoning, generation, and editing without separate encoders or VAEs. It suggests that multimodal understanding can transfer directly to complex visual planning and creation in an end-to-end architecture.

• Supports native resolutions up to 4K using spatially coherent patch reconstruction and curated generation and editing data.
• Uses specialized post-training experts for aesthetics, bilingual text rendering, infographics, and editing, consolidated through multi-expert on-policy distillation.
• Reported gains span image fidelity, text rendering, complex composition, multi-reference editing, instruction following, and subject and geometry preservation.
• The authors plan to open-source code for supervised fine-tuning, reinforcement learning, and on-policy distillation.

TL;DR - WORLDVIEW reveals that hidden prompt-revision layers in commercial text-to-image systems can causally introduce cultural stereotyping. This shows that bias audits must examine deployed generation pipelines—not just models or final images.

• WORLDVIEW contains 8,960 prompts spanning 15 languages and 31 language-context pairings.
• The audit covers prompt revision in DALL-E-3, Imagen-4, and GPT-Image-1.5.
• Compared with a no-context English baseline, US contexts were least marked, while non-Western and non-Anglophone contexts received heavier cultural marking.
• Revised prompts often compressed diverse contexts into narrow, stereotypical vocabularies; original-versus-revised prompt comparisons isolated revision as the cause.

TL;DR - OmniHallu is a unified framework and benchmark for detecting hallucinations across multimodal comprehension and generation tasks involving images, video, audio, and text. It matters because prior detectors are often limited to a single modality or task type.

• OmniHallu-Bench contains 10,000 samples with claim-level human annotations across six text-media conversion tasks.
• A multi-agent architecture decomposes outputs into atomic claims, uses modality-specific experts for verification, and aggregates their evidence through structured reasoning.
• A preference-optimized trainable verifier approximates the multi-agent decisions while reducing expert calls by 66% with minimal reported performance loss.
• Experiments identify a consistent modality-dependent performance gradient and provide fine-grained analysis of cross-modal hallucination patterns.

TL;DR - MindTopo is an 11,030-instance benchmark testing foundation models’ topological reasoning and closed-loop planning across continuity, separation, order, enclosure, and knots. Fourteen multimodal LLMs substantially trail humans, especially when planning requires preserving topology across actions.

• The benchmark spans 13 procedurally generated task types with controllable difficulty and evaluates both reasoning and agentic planning.
• Every tested multimodal LLM performed better on reasoning than planning; even the strongest remained far below observed human performance.
• Supervised fine-tuning and reinforcement learning improved Qwen3-VL-2B-Instruct’s reasoning more than its planning.
• Image- and video-generated observations preserved local cues and plausible endpoints but often violated environment dynamics or topology between transitions.

TL;DR - OmniKVQuant is a training-free KV-cache quantization framework for multimodal LLMs that accounts for differences across audio, video, and text representations. It enables 2-bit caches while substantially preserving performance, reducing memory demands without reconstructing a dense FP16 cache during decoding.

• Identifies temporal key drift and heterogeneous value geometry as key limitations of applying text-oriented TurboQuant to multimodal caches.
• Sets key quantization ranges over short input-stream windows and applies separate value rotations for each modality.
• Demonstrates 2-bit KV caches on Qwen2.5-Omni and Qwen3-Omni across seven audio-visual benchmarks.
• Provides a fused Triton decoding kernel that unpacks quantized values directly during attention.