237

Papers

10/10

Roadblocks Active

2

Connections

ContextCurator trains a small 7B policy model via reinforcement learning to decide what information an agent needs to keep, prune, or summarize — decoupling memory management from task execution. On real web navigation tasks, this raises success rates from 36.4% to 41.2% while cutting token use by 8.8%, and on multi-hop research tasks it cuts token consumption by a factor of 8 with improved accuracy. This matters because context overflow is one of the hardest practical blockers for deploying long-running agents: this is the first RL-trained system that handles it as a learned skill rather than a hard rule.

██████████ 0.9 long-context Preprint

ReadSaveConnections

By deploying the same frozen 8B model in three distinct roles — summarizer, agent, and corrector — within a single inference pipeline, this work roughly doubles task completion rates on multi-step tool-use benchmarks without any additional training. A scaffolded Qwen3-8B outperforms the much larger DeepSeek-Coder 33B on structured tasks, suggesting that architecture of execution matters as much as raw model size. This is a practical result: organizations constrained to small, locally-deployable models can recover substantial capability purely through inference design.

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

ReadSaveConnections

Instead of running one agent and hoping for the best, AggAgent runs multiple parallel trajectories and deploys a lightweight aggregation agent to synthesize them — inspecting and searching across solutions rather than just voting on final answers. This yields up to 5.3% absolute improvement on agentic benchmarks and 10.3% on deep research tasks, with overhead bounded by a single extra rollout. The key insight is that parallel test-time compute is cheap to scale, and aggregation is a learnable skill that extracts more signal from diversity than majority vote ever can.

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

ReadSaveConnections

PaperScope constructs 2,400 QA pairs across 25,495 AI papers and 11 sub-tasks to stress-test whether AI systems can genuinely reason across large bodies of scientific literature. Top systems including OpenAI Deep Research and Tongyi Deep Research achieve limited scores, with multi-hop reasoning over multi-modal evidence proving especially difficult. This benchmark matters because it exposes the gap between systems that retrieve well in isolation and systems that can synthesize evidence across documents the way a skilled researcher actually does.

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

ReadSaveConnections

GRIP embeds retrieval decisions directly into the autoregressive token stream: a model emits special control tokens to trigger retrieval, reformulate queries, or terminate search — all within a single decoding pass and without any external retrieval controller. This end-to-end coordination means the model learns when it actually needs outside information rather than always retrieving or never retrieving, which reduces both hallucination and wasted computation. The mechanism is important because it eliminates the architectural seam between generation and retrieval that currently forces RAG pipelines into brittle two-stage designs.

██████████ 0.9 hallucination-grounding Preprint

ReadSaveConnections

SWE-AGILE addresses the tension in software engineering agents between deep chain-of-thought reasoning (which bloats context) and shallow reactive agents (which miss edge cases) by maintaining only a sliding window of detailed reasoning and compressing older steps into concise digests. Trained on just 2,200 trajectories with a token-compression reward signal, it sets a new state-of-the-art for 7B–8B models on SWE-Bench-Verified. This shows that explicit context lifecycle management — not just bigger models — is the key lever for sustained reasoning in multi-turn coding tasks.

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

ReadSaveConnections

Developed in collaboration with 502 investment bankers from firms including Goldman Sachs and JPMorgan, this benchmark evaluates nine frontier LLMs on realistic end-to-end workflows producing Excel, PowerPoint, and PDF deliverables. The best model (GPT-5.4) fails nearly half of rubric criteria, and 0% of outputs were rated client-ready by practitioners. This hard number matters: it quantifies exactly how far current agents are from professional deployment thresholds in a high-stakes domain where errors have direct financial consequences.

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

ReadSaveConnections

Synthius-Mem organizes conversational memory into six cognitive domains inspired by human memory architecture, achieving 94.37% accuracy on the LoCoMo benchmark — surpassing both the previous best system (MemMachine at 91.69%) and human performance (87.9 F1). Adversarial robustness reaches 99.55%, a metric no competing system reports. This matters for deployed conversational agents, where persistent persona consistency and resistance to contradictory inputs are the difference between a useful assistant and an unreliable one.

██████████ 0.9 hallucination-grounding Preprint

ReadSaveConnections

Relax is a reinforcement learning training infrastructure that decouples data generation, model updates, and reward computation into independent asynchronous services, enabling fully async training to run 2x faster than co-located baselines on 30B multimodal models without sacrificing convergence. The architecture is built to handle text, image, video, and audio natively in a single stack, addressing the fragmented tooling problem that currently forces teams to maintain separate pipelines per modality. For organizations doing post-training at scale, this is a direct reduction in compute cost for RLHF and GRPO runs.

██████████ 0.8 efficiency-scaling Preprint

ReadSaveConnections

A neural network trained to count through physical robotic interaction reaches 96.8% accuracy using only 10% of the training data that a vision-only baseline requires — and the advantage persists even when visual-motor correspondences are randomized, indicating embodiment acts as a structural learning prior rather than a direct information source. The trained model spontaneously develops logarithmically-tuned number-selective units and a mental number line organization matching biological findings. This suggests that embodied interaction shapes representation in ways that transfer to abstract numerical reasoning, with implications for how much data efficient AI systems might need if grounded in physical experience.

██████████ 0.8 embodied-ai Preprint

ReadSaveConnections