cs.CV
2026/07
Depth foundation models generalize robustly zero-shot but are too heavy for embedded and mobile platforms; lightweight single-domain alternatives fail silently under domain shift. ZipDepth pairs an efficient reparameterizable encoder-decoder with multi-domain distillation from a foundation model. At 6.1M parameters it runs in real time from server GPUs to power-constrained devices, with the best zero-shot accuracy vs efficiency trade-off among lightweight models on five benchmarks (ECCV 2026).
Read more
cs.CL
2026/07
Existing benchmarks for proactive agents lean on sandboxes and single-turn evaluation, mixing capabilities in one task category so failure causes stay hidden. UniClawBench is the first capability-driven benchmark, built around five foundations (skill usage, exploration, long-context reasoning, multimodal understanding, cross-platform coordination). It evaluates 400 bilingual real-world tasks in live Docker containers with step-by-step checkpoints and simulated multi-turn feedback.
Read more
cs.AI
2026/07
Scientific ideas rarely start from a blank page: they inherit mechanisms, repair limitations, and recombine earlier work, like genomes. Existing benchmarks say little about whether AI can follow this inheritance. IG-Bench represents each paper as typed, evidence-grounded Idea Genome objects, with a GenomeDiff recording inheritance, mutation, loss, import, and insertion, across 10 domains. The strongest of 14 LLM systems reaches only 27.3% exact accuracy on lineage reasoning — a compositional bottleneck.
Read more
cs.AI
2026/07
A large-scale descriptive analysis of Syntea, an AI-based learning assistant in higher education, based on objective log data from 77,543 distance-learning students. Prior educational-chatbot research relied on small samples and self-reported surveys; large-scale evidence on actual usage was scarce. The study finds Syntea already embedded in the study routines of many learners, with usage differing across gender, age group, study cluster, degree, and study mode.
Read more
cs.AI
2026/07
Vision-Language Models have improved autonomous-driving tasks, but evaluating whether they reason reliably about safety-critical incidents remains hard. AUTOPILOT-VQA is an incident-centric visual question answering benchmark for dashcam video, built around real driving incidents and near-misses. It covers weather, traffic, road layout, signage, accident occurrence, impact location, and avoidability — moving beyond object recognition toward temporally grounded, safety-aware reasoning (AUTOPILOT CVPR 2026).
Read more
cs.AI
2026/07
Post-training quantization (8-bit down to 2-bit) is standard for deploying LLMs, yet evaluation leans on accuracy and perplexity. This work introduces correctness agreement, a decision-level metric measuring the overlap in correct predictions between a base model and its quantized variants. Behavioral divergence emerges under moderate quantization even when task performance appears preserved; query and key projections are consistently more sensitive than value and output projections.
Read more
cs.LG
2026/07
LLMs contain Super Weights: parameters whose removal degrades performance by orders of magnitude. This work shows the intuition "if they matter, train them" is wrong. Training them in isolation drops accuracy to random-guessing on OLMo-1B and OLMo-7B, while training equally many random positions instead beats baseline — the collapse comes from targeting Super Weights, not sparsity. Vanilla LoRA succeeds with 0.16% of parameters: importance does not imply trainability in isolation.
Read more
cs.AI
2026/07
In long-horizon tasks, decision-relevant state (requirements, environment facts, prior attempts, diagnoses, open subgoals) gets buried in or pushed beyond the context window — a failure mode the authors call behavioral state decay. A separate memory agent runs alongside an unmodified action agent, maintaining a structured memory bank and deciding whether to inject a memory-grounded reminder or stay silent. It lifts pass@1 by +8.3 pp on Terminal-Bench 2.0 and +6.8 pp on tau2-Bench.
Read more
cs.CL
2026/07
Reinforcement learning increasingly uses a judge LLM to score rubric criteria, acting as the reward model. Before trusting that signal, we must know how capable and biased the judge is. On citation quality in deep-research systems, the authors score 8 off-the-shelf judges over 1,248 human-reviewed decisions and find cheaper judges stay competitive. But at equal F1, judges differ sharply in pass-rate drift and false-positive/negative rates — directional bias that scalar F1 hides and an RL loop amplifies.
Read more
cs.SE
2026/07
Repository-level code generation must implement a target function amid cross-file dependencies and project conventions. Existing retrieval relies on lexical, structural, or semantic similarity, missing functions with similar procedural logic. ProjAgent adds procedural similarity as an explicit signal: it decomposes the target into steps, retrieves similar procedures per step, merges them with semantic retrieval, and repairs code via static analysis — reaching 41.14% Pass@1 on REPOCOD.
Read more
cs.SE
2026/06
Code LLMs need repository-level context to resolve imports, APIs, and conventions. Code2LoRA is a hypernetwork that generates repository-specific LoRA adapters, injecting that knowledge with zero inference-time token overhead. It offers a Static mode (snapshot → adapter) and an Evo mode updated per code diff.
Read more
cs.RO
2026/06
Manipulation alternates between low-risk transit (fast) and high-risk contact (slow, precise), yet existing Vision-Language-Action models (VLAs) inherit a single fixed speed from demonstrations. TempoVLA notes that the magnitude of each predicted action already governs speed, and controls execution speed via an explicit condition — combining a data-side variable-speed trajectory augmentation (VSTA) with model-side speed conditioning to control both acceleration and deceleration.
Read more
cs.CL
2026/06
As AI writing assistants spread, documents are increasingly the product of progressive human–AI co-editing rather than purely human or AI. OpAI-Bench studies human-to-AI transformation at document, sentence, token, and span granularities — and finds that mixed-authorship "intermediate" versions are often harder to detect than fully human or heavily AI-edited endpoints (a non-monotonic pattern).
Read more
cs.LG
2026/06
Standard RNN training (BPTT) is sequential in time, hard to parallelize, and suffers vanishing/exploding gradients on long ranges. SMT reduces RNN training to supervised learning on one-step memory-transition labels (m_t, x_{t+1})→m_{t+1}, sidestepping recurrent credit propagation entirely — enabling time-parallel training with an O(1) gradient path between any two tokens. It beats BPTT on language and pixel-sequence modeling (MIT, Isola lab).
Read more
cs.CV
2026/06
Vision-Language Models (VLMs) tend to confine reasoning to observed images and text, struggling with unobserved layouts and alternative viewpoints. Astra is a "thinking with imagination" framework where a VLM actively queries a world simulator for imagined novel-view evidence during reasoning. It couples an RL-trained policy with a Bagel-based world model and improves spatial-reasoning benchmarks.
Read more
cs.LG
2026/06
RL fine-tuning of reasoning models (e.g., GRPO) can only verify and reward the final answer after the chain-of-thought (CoT) is complete — a delayed-reward problem that is Monte-Carlo-like and high-variance. RREDCoT redistributes reward (credit assignment) to the CoT segments that mattered, approximating the optimal redistribution using the model itself without extra generation (from LSTM creator Hochreiter's group, JKU).
Read more
cs.AI
2026/06
When LLM agents tackle long-horizon tasks like ML engineering, inter-branch information isolation, memoryless search, and a lack of hierarchical control hamper long-horizon optimization. MLEvolve is a self-evolving multi-agent framework that enables cross-branch information flow and reuses accumulated experience. It reaches SOTA on MLE-Bench in half the time budget and beats AlphaEvolve on math optimization.
Read more
cs.AI
2026/06
An agentic framework for formal theorem proving in Lean 4 that generates and refines a "blueprint" — a dependency graph of definitions and lemmas. A tool-equipped Lean prover closes each lemma node in parallel, and failures drive blueprint refinement, avoiding the dead-end loops of recursive decomposition. On an open-weight backbone it reaches 99.2% on MiniF2F and 75.6% on PutnamBench (88.8% with a natural-language proof) — SOTA-class for an open-source pipeline.
Read more
cs.CL
2026/06
Long-context inference is bottlenecked by decoding efficiency, especially for reasoning models that emit long chains of thought. Existing sparse attention faces an efficiency-quality trade-off. CLSA, built on KV-sharing (YOCO), shares not just the KV cache but the routing index across layers — computing top-k selection once and reusing it. At 128K context it reaches up to 7.6x decoding speedup and 17.1x overall throughput.
Read more
cs.CR
2026/06
As autonomous LLM agents hold real credentials and operate infrastructure, operators lack a standard way to say a resource is off-limits. The Recuse Signal is a lightweight in-band deny signal (over an SSH banner or a PostgreSQL NOTICE) asking an automated agent to voluntarily withdraw — a robots.txt-like cooperative control, not a security boundary. In a pilot, the signal induced 100% recusal versus 100% task completion without it.
Read more
