288

Papers

13/13

Roadblocks Active

3

Connections

MEME introduces a benchmark that tests AI memory systems across tasks requiring tracking of multiple entities and facts that change over time. The headline finding is that all six evaluated memory systems collapse on dependency reasoning — where understanding one fact requires correctly chaining it to another — achieving just 1–3% accuracy on these subtasks despite adequate performance on simpler retrieval. This matters because dependency reasoning is exactly what long-horizon AI assistants need, meaning current memory designs are not fit for practical multi-session use.

██████████ 0.9 reasoning-reliability Preprint

ReadSaveConnections

Training robot-control policies with reinforcement learning normally requires building a simulation model specific to each task, which is expensive. This paper (RAW-Dream) shows that a world model pre-trained on diverse, task-free robot behaviors can substitute for task-specific simulation, allowing reinforcement learning using zero-shot binary rewards from a frozen off-the-shelf vision-language model. A dual-noise verification mechanism filters out the hallucinated rollouts that world models inevitably produce, and the approach generalises zero-shot across multiple held-out task suites.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

GUI agents that re-interpret the screen with a large language model at every step become fragile on long tasks because small errors accumulate. EAM replaces step-by-step generation with a knowledge graph built offline via state-aware search, compressing multi-step routines into retrievable action groups that the agent executes rather than re-derives. Evaluation on Android and mobile benchmarks shows improved reliability over LLM-only baselines, with theoretical bounds on how accurately the system can recover planned paths.

██████████ 0.9 agent-tool-use Preprint

ReadSaveConnections

Failure analysis of GPT-5.4 controlling a Windows desktop reveals that computer-use agents fail disproportionately on rare interaction types — complex drags, precision clicks, unusual shortcuts — while handling common ones well. The root cause identified is data scarcity: these low-frequency interactions simply appear too rarely in training data. The authors build a renderer-based pipeline generating 50 million synthetic training examples covering the long tail, and fine-tuning on this corpus measurably improves performance on the CUActSpot benchmark for hard interactions.

██████████ 0.9 agent-tool-use Preprint

ReadSaveConnections

Most vision AI models are built separately for understanding images versus generating them, requiring two different systems in practice. SenseNova-U1 trains a single architecture (NEO-unify) on both tasks jointly using autoregressive language prediction and image-space flow matching as simultaneous objectives, without relying on pretrained vision encoders. Released weights on HuggingFace let the community test whether unified models can match specialist understanding-only systems while also generating and editing images.

██████████ 0.9 multimodal-understanding Preprint

ReadSaveConnections

MM-OptBench tests whether AI can translate real-world optimization problems — presented as text, tables, and charts — into correct mathematical programs verified by exact solvers. The best general-purpose models reach only 52% accuracy, performance degrades from 43% on easy instances to 16% on hard ones, and all three math-specialized models score exactly 0 out of 780 problems. This last result is surprising: models designed for mathematics fail completely when the problem requires integrating visual and structured information.

██████████ 0.8 multimodal-understanding Preprint

ReadSaveConnections

SAGE builds a graph-structured memory that improves itself over multiple rounds: a memory-writer agent constructs the graph from conversation history, while a graph foundation model reads it and feeds back signals to improve how future information is stored. After two self-evolution rounds the system achieves the best average rank on multi-hop QA benchmarks, and zero-shot transfer to open-domain retrieval reaches 82.5% Recall@2 on Natural Questions. The self-improvement loop is the key mechanism — memory quality rises without manual curation or additional labelled data.

██████████ 0.8 hallucination-grounding Preprint

ReadSaveConnections

Applying reinforcement learning to models that jointly generate text and images is unstable because complex multimodal outputs are hard to score automatically. AlphaGRPO addresses this with a Decompositional Verifiable Reward that breaks each generation request into atomic yes/no questions (e.g., 'Is the object red?', 'Is the object on the left?') scored by a verifier model, giving a stable and interpretable training signal. Models trained this way learn to diagnose and correct errors in their own generated images — a self-reflective capability not achieved without this reward structure.

██████████ 0.8 reasoning-reliability Preprint

ReadSaveConnections

LLM-based service agents routinely violate policy constraints, hallucinate tool calls, or misread user intent because they maintain task state implicitly in their context window. The NOD architecture externalises state into a shared Global State structure visible to all agents, separates action generation (Operator) from verification (Director), and only executes high-stakes actions after the Director independently confirms they are safe and intentional. This structural oversight reduces all three failure modes without retraining the underlying language model.

██████████ 0.8 agent-tool-use Preprint

ReadSaveConnections

LongMemEval-V2 benchmarks whether web agents can accumulate domain-specific knowledge across tasks and reuse it like an experienced colleague rather than starting fresh each time. The best-performing approach stores past agent trajectories as files and retrieves them via a coding sub-agent, reaching 72.5% accuracy versus 48.5% for retrieval-augmented generation and 69.3% for a plain coding agent baseline. The benchmark's five-skill taxonomy — static state recall, dynamic state tracking, workflow knowledge, environment quirks, and premise awareness — provides a more diagnostic framework than single accuracy scores for identifying where agent memory breaks down.

██████████ 0.8 agent-tool-use Preprint

ReadSaveConnections