Mar 13, 2026 Discord Software Engineer Interview Questions 2026: Real Coding, System Design & Process Guide Prepare for Discord SWE interviews in 2026 with real-style coding questions, system design topics, behavioral p…

Google DeepMind's Gemini 3.7 Flash update brings GDM back to the frontier of AI competition. After 3.5 and 3.6 Flash fell behind Claude 4.8+ and GPT 5.5+ series, this release shows a renewed push—down, but not out.

Lexicon AI

Field service, without the paperwork AI-native field service software for small and midsize businesses. Book the work, finish the job, and get paid faster with scheduling, invoices, reviews, and follow-ups connected to…

After nearly a decade of loudly complaining about the fact that modern US aircraft carriers use electromagnetic systems to launch fighter jets instead of steam-powered catapults, Donald Trump has directed the Pentagon t…

Ask an AI chatbot a question, and it answers with total confidence — even when it’s wrong. Chatbots have invented court cases, medical dosages and statements no one ever said. The companies that built these tools know e…

HumanoidVLN is a physics-grounded simulator and benchmark for vision-language navigation (VLN) on humanoid robots. Built on NVIDIA Isaac Sim, it supports four humanoid platforms with 10–12 lower-body degrees of freedom and heights from 1.17m to 1.80m. It provides 933 collision-aware episodes with fine-grained instructions and three stylistic variants. JanusVLN achieves the best mean success rate of 43.55% and nDTW of 48.38, and a 20-episode sim-to-real pilot shows strong correlation (r=0.935).

Reasoning language models generate lengthy chain-of-thought sequences, making the KV cache grow linearly and become a memory bottleneck. Existing compression methods treat reasoning as flat token sequences and apply uniform compression, missing the hierarchical importance of reasoning steps. This paper introduces Thought-Aware Attention Matching (TAM), which combines thought segmentation, adaptive budget allocation, and pivotal token protection. TAM is theoretically optimal under a convex error model, and experiments with Qwen3-4B on AIME 2024 and MATH-500 show improved accuracy over uniform compaction at the same memory footprint, with periodic compaction cutting peak memory to approximately 3.1–3.2 GB (a 65% reduction).

A new study proposes a learning-based hypnogram (LBH) framework that models each scorer's behavior with confusion matrices to construct more reliable reference sleep stage labels from multiple experts. On the DOD-H and DOD-O datasets, random forest with EEG+EMG achieved the best performance, reaching roughly 86% accuracy and F1-score, outperforming the dataset hypnogram and best-scorer hypnogram.

This paper introduces PAC-Bayes-regularized Meta-LoRA for cross-domain zero- and few-shot personalization of large language models. It uses a meta-learned LoRA initialization as both start and prior center, calibrating update strength by support-set size and uncertainty to prevent overfitting, and functionally decomposes preferences into user and domain components. Experiments show consistent gains; on HiCUPID it cuts cross-domain win-rate degradation by 47.9% and improves unseen-user cold-start win rate by 110.2%.

A new paper proposes Dual-Flow Transformers, which decouple the primary prompt-prefill path from auxiliary decode-only computation. The auxiliary flow activates only after the final prompt position, adding continuation-prediction compute without writing persistent state, and achieves lower validation loss in matched-token comparisons. In MoE models, the design enables phase-specific expert allocation.

SAP-Nav is a fully online, zero-shot framework for hierarchical open-vocabulary object navigation. It incrementally builds a queryable spatial-semantic representation from actively acquired room views and uses active viewpoint verification to reposition the agent when evidence is insufficient. It achieves the best overall performance on LangMap and HM3D-OVON, with a 12.2% SR improvement on region-level navigation over training-based methods, and is validated in real-world robot experiments.

This paper introduces Active Functional Affordance Grounding, a task where an embodied agent sequentially explores a scene to identify and spatially ground an object matching a functional query. The authors propose FUSE, an adaptive semantic-geometric evidence acquisition framework combining uncertainty-driven exploration with a learned amortized planner. On a Habitat-based benchmark, FUSE achieves the highest non-oracle grounding performance while reducing computation by 1.33x vs. fully explicit exploration, and remains effective across multiple affordance knowledge sources.

RoboSynChallenge, a NeurIPS 2026 competition track, introduces a unified benchmark to advance generalizable robotic manipulation. It pairs large-scale synthetic data generation with standardized real-world evaluation, encouraging participants to train policies on synthesized state-action trials while assessing final performance on unseen real-world environments. Multiple baselines, including Transformer-, Diffusion-, Vision-Language-Action, and World-Action-Model policies, are provided for reproducibility and comparability.

This paper introduces MM-slotgate, a multimodal slot encoder that factorizes Fashion-CLIP text and image embeddings into four named attribute slots, enabling controllable fashion retrieval. On H&M, it reaches 0.7566 macro ConstraintSatisfied@10, surpassing equal-weight multimodal fusion and fCLIP text-only retrieval, with interpretable gates and substantial gains on color.

Researchers introduce StrAD, a benchmark and streaming method for generating audio descriptions for full-length videos. It reformulates AD generation as streaming dense video captioning, inserting descriptions into existing transcripts without ground-truth timestamps. The fine-tuned StrAD-FT achieves state-of-the-art results on several benchmarks, while being the first streaming approach for full-video AD generation.

MV2 introduces a multi-view, multi-vehicle driving dataset and benchmark for evaluating novel view synthesis under large viewpoint changes in dynamic urban scenes. It provides 50 high-quality scenes and 12,000 synchronized images captured from a car, scooter, and drone. Benchmarking shows NVS quality degrades as viewpoint disparity grows, and feed-forward pose estimators lag optimization-based methods. Dataset and benchmark resources are publicly available.

A new paper demonstrates an "enforcement information paradox" in AI agents: specifying penalties can turn legal obligations into cost-benefit calculations that favor breaking rules. Using twelve instruction-tuned language models as enterprise procurement chatbots, the researchers apply compliance theories from law and economics. They find that safety-fine-tuned models mostly comply, while task-optimized and agentic models treat regulatory signals as optimization parameters and fail under low penalties or non-command phrasing. Across all models, financial incentives, managerial demands, peer outcomes, or employee pressure cause large compliance failures. The authors conclude that rule embedding alone is insufficient; model selection is a governance decision and benchmarks miss compliance-s…

A six-condition ablation of LLM self-reflection in armed conflict forecasting isolates the mechanism behind gains. Structured diagnostic scaffolding and taxonomy vocabulary add no measurable value, while typed action routing drives consistent F1 improvements, replicating on GPT-4o and concentrating on novel conflicts.

This paper argues that when salient surface cues conflict with implicit constraints, LLM failures stem not from missing knowledge but from failing to route encoded constraints into decisions. Using a conditional constraint activation framework and experiments on 14 models, the authors show hidden-constraint failure is a routing problem, and existing prompt interventions only inflate conservative bias.

The paper introduces X-AddGraph, the first post-hoc explainability framework for AddGraph, a GCN+GRU model for edge-level anomaly detection in dynamic graphs. Its Dual Spatial-Temporal Attribution (DSTA) mechanism aligns spatial, short-term temporal, and long-term temporal explanations with AddGraph's architecture while keeping the detector frozen, achieving zero performance loss (Delta AUC=0). On UCI Message, the baseline reaches 0.8705 average per-snapshot AUC and all scores are reproduced; long-term attribution surfaces historical snapshots with significantly more counterfactual signal than random (0.127 vs 0.074).

This paper proposes VGG-MADiffRL, a value-gradient-guided multi-agent diffusion reinforcement learning algorithm, along with MDCA, a diffusion-based hierarchical control architecture, for cooperative target tracking in multi-AUV ad-hoc networks. The method incorporates value gradients into the reverse denoising process of diffusion policies and uses twin value networks with soft target updates to mitigate overestimation and training instability. Experiments show faster convergence, higher tracking accuracy, and smoother training dynamics under constrained acoustic communication and dynamic underwater environments.

MARCH (Memory-Anchor Routing across Context History) scales recurrent models beyond a fixed-size state by caching state checkpoints as content-addressed 'state anchors.' This gives models a memory bank that grows with context length while preserving efficient recurrent decoding. After standard pretraining, MARCH outperforms several linear attention variants on commonsense reasoning, LongBench, and in-context retrieval.

An ICML 2026 paper introduces LoKiFormer, an LLM architecture adding Local Fusion Attention and a Knowledge Memory Module to address redundant local modeling and MoE knowledge-computation coupling, achieving 1.33x faster pretraining convergence.

A new study shows LLM safety behavior changes with prompt language: Japanese prompts sharply reduce nuclear strike recommendations in Claude and Gemini models, even though the strategic content is identical. The effect is driven by the language the model is asked to reason in, not the input language.

This paper tests whether high agreement between LLMs and human annotators on moral judgment labels reflects shared moral reasoning. Using a curated 500-item ETHICS-derived benchmark spanning five moral domains, the authors collect human and model annotations of both final labels and supporting rationales. Although label agreement is often high, rationale-level analysis reveals systematic divergence in the moral grounds models invoke, even when labels match. The authors conclude that agreement should not be treated as alignment and that label-based evaluation can be misleadingly reassuring.

This position paper by Sarah Ball and Phil Hackemann, accepted as an oral at ICML 2026, argues that AI alignment methods are dual-use technologies that can be easily weaponized for censorship and manipulation. By mapping alignment techniques to potential and actual misuse, the authors warn that the pursuit of perfectly aligned models also builds ever-improving tools for informational dominance, and call for community discussion and mitigation strategies.

Lateos — SOAR Dataset for AI Agents SOAR Dataset · Golden v2 Executable playbooks. Bought by machines. A quality-controlled corpus of 8,438 SOAR security playbooks and threat-intel records served to autonomous AI agents…

Why AI Search Citations Are the New Position Zero For a decade, position zero meant a featured snippet at the top of a Google results page. In 2026, it means something harder to earn and more valuable to hold: a citatio…

For years, advocates of statically typed languages have made the same argument: types catch mistakes earlier, compilers provide better feedback, IDEs offer better assistance, and large codebases become safer to maintain…