283

Papers

10/10

Roadblocks Active

3

Connections

Gyan replaces the transformer architecture with a rule-based neuro-symbolic pipeline that builds structured meaning graphs using linguistic and rhetorical theory, explicitly separating language understanding from statistical pattern matching. The system claims state-of-the-art results on three public benchmarks including MS Marco and MMLU-Medicine without learned statistical weights, which would matter enormously for AI reliability — but the code is proprietary and two of five evaluation datasets are inaccessible, making independent verification impossible. If the claims hold up under scrutiny, this architecture offers a path to AI systems whose reasoning steps can be audited rather than inferred.

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

ReadSaveConnections

LongSeeker trains an agent to actively manage its own working memory during long multi-step searches using five atomic operations — skip irrelevant content, compress resolved information, roll back dead ends, clip important evidence, or delete to control size — rather than passively accumulating everything in a growing context window. The model is fine-tuned on 30 billion parameters and evaluated on four benchmarks including GAIA, with code and weights publicly released. This matters because context bloat is one of the primary reasons long-horizon agents hallucinate or lose track of task state, and a learnable context management layer addresses the root cause rather than just extending context limits.

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

ReadSaveConnections

DTap is an open red-teaming platform that simulates 50+ real-world agentic environments — including Gmail, PayPal, and Slack — and uses an autonomous attack agent to find vulnerabilities across five injection vectors (prompt, tool, skill, environment, and combinations). Testing current agents built on GPT, Gemini, Claude, and DeepSeek backbones, the platform documents that AI agents can be reliably manipulated into leaking API keys, deleting user data, and initiating unauthorized financial transactions. This is a concrete empirical demonstration that deployment-ready agents across all major model families share systematic exploitable weaknesses, not just theoretical concerns.

██████████ 0.9 alignment-safety Preprint

ReadSaveConnections

This system chains a small 4-billion-parameter language model with a classical automated theorem prover (Prover9): the LLM translates natural language syllogisms into formal logic, and the prover handles the deduction step deterministically. The result outperforms zero-shot LLM baselines at the same parameter scale on formal reasoning tasks, while largely avoiding the 'content effect' — where models reach different conclusions depending on whether premises are semantically plausible rather than logically valid. The approach demonstrates that hybrid symbolic-neural pipelines can give small models formal reasoning capabilities that scale poorly through pure statistical training.

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

ReadSaveConnections

OpenSearch-VL provides a fully open training recipe — datasets, code, model weights — for building multimodal agents that can search the web, enhance images, and parse documents across multiple turns of interaction, achieving over 10-point average gains across seven benchmarks. The key training innovation is a 'fatal-aware' reinforcement learning algorithm that prevents the agent from being penalized for actions taken after a tool has already failed, making learning more stable when tool pipelines break mid-task. Matching several proprietary commercial model benchmarks using open components is a practical milestone for reproducible agentic AI research.

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

ReadSaveConnections

ConsisVLA-4D adds three modules to a vision-language-action robot model that align visual features across multiple camera viewpoints and reason about how objects move over time — using only standard RGB cameras without depth sensors. On the LIBERO simulation benchmark this yields a 21.6% performance improvement over the baseline OpenVLA, and 41.5% on real-world robot tasks, with a 2.3× inference speedup. The approach matters because most robot learning systems either require expensive depth sensors or struggle with spatial consistency when objects are viewed from different angles during manipulation.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

LMMP separates the planning and execution stages of Earth observation agents: a planner that combines satellite image features with task semantics selects which tools to call from a library of 100+ specialized remote sensing tools, while a frozen executor carries out the plan. The design injects domain expert knowledge through a structured 'Meta Task Library' that constrains the planner to physically plausible tool sequences, reducing invalid tool calls. Tool-calling accuracy improves significantly over end-to-end baselines, demonstrating that domain knowledge injection at the planning layer is more efficient than fine-tuning generalist models on specialized tasks.

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

ReadSaveConnections

This paper tests whether a coding agent can learn game rules by writing and iteratively refining Python programs that predict game outcomes, using the new ARC-AGI-3 benchmark which requires understanding novel games from scratch. The agent fully solved 7 of 25 public games using no game-specific code, relying instead on a verifier that checks whether the program correctly predicts observed game states — with simpler programs preferred as a proxy for genuine rule understanding. The result is modest but significant as a proof of concept that executable world models with built-in verification can generalize to genuinely novel structured reasoning tasks without task-specific engineering.

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

ReadSaveConnections

Uno-Orchestra learns a single policy that simultaneously decides whether to break a task into subtasks, which specialized worker agent to assign each part to, and how much compute budget to allocate — rather than fixing these decisions by hand or using separate models for each choice. Across a 13-benchmark suite it achieves 77% pass rate, roughly 16 percentage points above the strongest fixed-workflow baseline, while reducing per-query inference cost by approximately an order of magnitude. The practical implication is that intelligent routing of tasks to appropriate agents is more cost-effective than always invoking the most capable (and expensive) model.

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

ReadSaveConnections

VTAgent investigates why video question-answering models underperform on text-heavy videos (e.g., reading signs, whiteboards, documents in video) and finds through oracle analysis that the bottleneck is locating the right frame, not reasoning once the frame is found. The system uses a question-guided agent to first identify and anchor the relevant keyframe before answering, improving accuracy by 12 points and ANLS by 11 points across benchmarks with supervised fine-tuning and reinforcement learning. This reframing — from 'how do we reason over video' to 'how do we find the evidence first' — provides a cleaner decomposition of the video QA problem.

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

ReadSaveConnections