Saral Research Paper

Brute force compute isn't the future of AI research—smart recursive architecture is.

When traditional AI agents handle complex deep research, they get buried under endless context logs and lose track of verified facts.

AREX (Recursively Self-Improving Deep Research Agent) solves this using a bi-level loop system:
- Inner Loop: Gathers evidence and tests hypotheses.
- Outer Loop: Acts as a strict supervisor, pruning useless context via Autonomous Context Updating (ACU).

The Result? AREX-Base scores 82% on wide search benchmarks, beating 397B+ parameter models (74%) while using a fraction of the compute.

Check out our latest breakdown video to see how step-aware RL is changing agentic workflows!

23 hours ago | [YT] | 0

Saral Research Paper

Is your team relying on temporary, AI-generated scripts for data pipelines?

Most AI coding agents output disposable code that breaks due to hallucinated dependencies and missing schema validations.

The NL2Pipeline Gap is real—but structured solutions are here. By pairing Model Context Protocol (MCP) with synchronized visual DAG editors, DataFlow-Harness slashes pipeline execution costs by 72.5% and cuts generation latency in half while boosting accuracy to 93.3%.

Check out our latest explainer video to see how it works!

3 days ago | [YT] | 0

Saral Research Paper

Did you know AI agents no longer need to rely solely on flat text documentation or raw code to master new software?

Humans learn complex software like Blender or Excel by watching video tutorials—taking in timing, visual cues, and spatial layouts. With Resource2Skill, Microsoft Research has brought that same multimodal learning process to AI!

By parsing open online tutorials through Vision LLMs, Resource2Skill distills raw video content into executable procedural knowledge stored inside a structured Skill Wiki. In real-world tests across 7 software domains (including UE5 and Web Design), this approach boosted AI agent performance by +11.9%!

What software task would you want an AI agent to learn for you next? Let us know below!

5 days ago | [YT] | 0

Saral Research Paper

Stop burning thousands of dollars on commercial API calls for Knowledge Graph extraction!

The Language vs. World Knowledge Hypothesis proves that while world knowledge scales with model size, core language comprehension plateaus early. Since RAG supplies the context, a compact 7B model can perform entity and relationship extraction just as accurately as 70B+ behemoths.

RAGU (Meno-Lite 0.1) delivers 82.4% evidence recall on medical benchmarks while running entirely on a single consumer GPU.

1 week ago | [YT] | 0

Saral Research Paper

Is your AI training hitting the GPU memory wall during RL post-training?

LongStraw proves that scaling to 2M+ tokens isn't just about adding GPUs - it’s about dynamic memory lifetime management! By using Serial Replay, it ran 4.4M+ tokens on just 8 GPUs.

What's holding your models back?

1 week ago | [YT] | 0

Saral Research Paper

Ever wondered why AI coding agents struggle to update complex codebases? It’s called the Behavior Localization bottleneck.

Discover how the Harness Handbook uses automated 3-tier mapping to boost AI agent efficiency by up to +18.9%! Watch our latest deep dive

1 week ago | [YT] | 0

Saral Research Paper

Did you know 55% of AI video benchmark questions can be solved WITHOUT watching the video?

The Video-Oasis paper reveals how Video LLMs use text shortcuts & static frames to 'cheat' tests. When blocked, scores crash to 25%. Are we training real video understanding or just test takers?

2 weeks ago | [YT] | 0

Saral Research Paper

Stop wasting money on LLM API calls for small daily coding tasks!

Program-as-Weights (PAW) compiles English descriptions into tiny, local LoRA weights. Result: A 0.6B local model running at 31.6 tok/s on an M3 Mac beat a 32B model in accuracy—at $0 token cost.

Is the future of AI tool-building local?

2 weeks ago | [YT] | 0

Saral Research Paper

Are per-repo Docker setups becoming obsolete for AI coding agents? 🚀

Our latest analysis breaks down the Dockerless Verifier - an environment-free pipeline achieving a 62% resolve rate on SWE-bench Verified and +14.3 AUC over DeepSWE without running containerized test suites.

2 weeks ago | [YT] | 0

Saral Research Paper

Trillion-parameter AI models consume massive energy, but are they always necessary?

Shanghai AI Lab’s Agents-A1 proves a 35B model can beat trillion-param giants by scaling Agent Horizon (45K token reasoning paths) instead of parameters. GAIA score: 96.0!

3 weeks ago | [YT] | 0