275

Papers

10/10

Roadblocks Active

0

Connections

The paper trains a 7-billion-parameter chemistry model (OmniChem) using a pipeline where multiple AI agents compete and cooperate — borrowing from game theory — to generate and filter high-quality training data, reportedly cutting hallucinations by 79% compared to the base model. The result is that a small specialist model reaches parity with GPT-4o mini on chemistry benchmarks, which matters because it suggests domain-specific hallucination is partly a data-quality problem, not just a model-size problem. The main caveat: the key benchmark (ThChem) is proprietary and not publicly available, making independent verification impossible for now.

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

ReadSaveConnections

UniClawBench tests AI agents on 400 real-world tasks running inside live Docker containers — not sandboxes — across five capability dimensions including cross-platform coordination and multimodal understanding. A key design choice is a hidden 'supervisor' agent that grades performance without leaking rubric details to the agent being tested, making it harder to game. This matters because most existing agent benchmarks use controlled, static environments that don't reflect the messy, dynamic conditions agents face in actual deployment.

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

ReadSaveConnections

WebSwarm breaks down complex web research tasks by recursively spawning specialized sub-agents that each handle a narrow search goal, then combining results — similar to how a research team divides a large investigation. The system outperforms single-agent and fixed multi-agent baselines on benchmarks covering deep (multi-hop) and wide (broad survey) search tasks. The practical implication is that recursive delegation, rather than a single powerful agent, may be the right architecture for open-ended information retrieval.

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

ReadSaveConnections

CausalDS builds a benchmark that tests whether AI agents can do causal reasoning — not just pattern matching — in realistic data-science workflows, spanning all three levels of Pearl's causal hierarchy (observation, intervention, counterfactual). Unlike prior causal benchmarks that use toy symbolic problems, this one requires agents to chain together data-analysis tools sequentially because observations are noisy and incomplete. This distinction matters because passing current benchmarks does not tell us whether an agent actually understands cause and effect or is just correlating inputs to outputs.

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

ReadSaveConnections

This paper tackles a specific problem in long multimodal conversations: as the dialogue grows, the number of image tokens explodes and the model loses track of which image was referenced when. The fix is an external 'Episodic Visual Memory' that stores structured summaries of past images and retrieves only what is relevant, rather than feeding everything into context. The 8-billion-parameter version achieves 91.4% retrieval accuracy over 20-turn sessions — beating 32-billion-parameter baselines by 8 percentage points — while cutting per-turn inference time nearly in half.

██████████ 0.8 long-context Preprint

ReadSaveConnections

The paper pits four frontier AI models (Gemini, Claude, Codex, Grok) against each other on coding problems and uses execution-based verification — actually running unit tests — to rank which teacher's solutions are correct, then feeds only verified solutions to a student model. A counterintuitive finding is that simply fine-tuning a competent student model (Qwen2.5-Coder at 7B and 32B) on these verified solutions makes it worse, dropping competition accuracy from 5.9% to 2.9%. This suggests that imitation of even correct solutions can suppress the flexible reasoning these models already learned.

██████████ 0.8 data-quality-curation Preprint

ReadSaveConnections

Rather than retraining large Vision-Language-Action robot models when they encounter new situations, Harness VLA wraps a frozen model inside an agent framework that delegates contact-heavy manipulation steps to the VLA and simpler transport or navigation steps to hand-coded routines. A memory module stores traces of successful runs and failure patterns, allowing the system to adapt to shifted spatial layouts or renamed goals without any gradient updates. The approach is practically significant because retraining large VLA models is expensive, and keeping the core model frozen reduces that cost to inference-time orchestration.

██████████ 0.8 embodied-ai Preprint

ReadSaveConnections

HCC-STAR is a specialized language model for liver cancer treatment recommendations, trained on roughly 30,000 cases and validated against 6,668 patients across 12 hospitals. It outperforms general-purpose models (GPT-5, Gemini-2.5 Pro) and clinical guidelines on treatment accuracy, and a hypothetical survival analysis suggests its recommendations correlate with 51-month median survival versus 29 months for standard guidelines. The study demonstrates that domain-specific training with reinforcement-learning-based reasoning can produce clinically meaningful improvements, though the validation cohort and model weights are not publicly released, limiting independent scrutiny.

██████████ 0.8 alignment-safety Preprint

ReadSaveConnections

The benchmark tests six vision-language models on progressively harder food-related tasks: identifying ingredients, estimating gram-level weights from photos, and giving dietary advice to patients with conditions like diabetes. Models perform well at naming dishes but fail badly at weight estimation and — more dangerously — frequently recommend unsafe foods to high-risk patients. The paper labels this a 'Semantic-Physical Gap': models understand visual categories well but cannot reason reliably about physical quantities or downstream health consequences, which is a concrete safety failure mode.

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

ReadSaveConnections

The paper identifies a structural collapse problem in a widely used technique for building symbolic world models: when language-learning gradients flow into a discrete symbol layer (Gumbel-softmax bottleneck), the model collapses to using only 2 out of 64 symbols in 4 out of 5 training runs. Five different engineering fixes all prevent collapse but fail to achieve semantic grounding (below 9.2% accuracy), suggesting the problem is architectural rather than a tuning issue. The proposed fix — blocking language gradients from reaching the symbol layer — resolves both failure modes simultaneously, with implications for any system that tries to combine symbolic and neural representations.

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

ReadSaveConnections