Trains LLMs with reinforcement learning using a surface chrF reward so models learn to extract and apply linguistic signals from rich context for translating completely unseen languages. Demonstrates better zero-shot translation than in-context learning or supervised fine-tuning, framing outcome-based RL as a meta-skill for language learning from context.
AI Agent Papers2026
Measures how coding agents explore repositories by asking them to return a ranked, line-level list of code regions relevant to an issue under a fixed line budget. Covers 848 issues across 203 repos and 10 languages; evaluates coverage, ranking, and context-efficiency to isolate exploration quality.
AI Video Papers2026
Decouples perception and reasoning for hours-long videos by streaming inputs into a three-tier Hierarchical Graph Memory and using an agentic Observation–Reason–Action retrieval loop; reduces reasoning context to ~2% of full video while improving benchmark accuracy.
Removes the subspace of frequent, uninformative tokens that LLMs inject into text embeddings via the model's unembedding matrix. EmbedFilter is a lightweight linear transform that refines LLM-derived embeddings to improve zero‑shot semantic retrieval, enable dimensionality reduction, and speed up indexing; code on GitHub.
Provides a comprehensive benchmark for instruction-based audio editing across seven audio modalities and eight operation types, with 2,000 high-fidelity samples and a rubric that decomposes tasks into 17,741 verifiable criteria for multi-dimensional evaluation.
Simulates egocentric, embodied human–world interactions and enables customizable, self-evolving local scenes by defining anchor views and text-driven evolution. Uses exogenous viewpoints and full-body motion supervision to improve spatial grounding and interaction consistency.
Analyzes the parameter-space geometry of on-policy distillation (OPD) for LLM training, showing OPD updates affect fewer weights, avoid principal directions, and rapidly lock into a low-dimensional update subspace. Compares OPD with supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR) and studies implications for optimization and objective mixing.
AI Agent Papers2026
Benchmark for long-horizon computer-use agents that must orchestrate GUI, CLI, and code operations within single trajectories across 114 real-world tasks. Evaluated on a real Ubuntu desktop and paired with a trajectory-aware judge that inspects deliverables, artifacts, and action traces—revealing a top PassRate of ~41.2%.
Models visual preference as distributions over rubric scores and introduces Z-Reward, a teacher–student framework that decouples reasoning-heavy judgment (teacher trained with GDSO) from efficient deployment (student via RISD). Demonstrates higher human-preference accuracy and works as a differentiable reward for text-to-image optimization.
Standardizes representation-level evaluation for tabular encoders by exporting row-, column-, and table-level embeddings and probing them with shared lightweight heads across three suites (TRL-CTbench, TRL-Rbench, TRL-DLTE). Supplies curated benchmark assets and task rewrites (50 OpenML tables, 123 targets, a 47,772-table DLTE lake) to enable fair cross-paradigm comparison.
Generates outcome-specific, dialectical rationales with an LLM and derives continuous, calibrated risk scores for irregularly sampled medical time series—mitigating risk polarization. Reports +3.3% average AUPRC and 81% reduction in calibration error across three benchmarks; code released.
AI Agent Papers2026
A benchmark that evaluates interactive spatial reasoning for multimodal agents in realistic tasks. It unifies eight heterogeneous simulators under a simulator-agnostic protocol, provides 760 human-annotated tasks with vision-only partial observability, and uses text-based actions plus terminal-state verification to measure task success.