283

Papers

10/10

Roadblocks Active

0

Connections

OmniAgent treats video understanding as a decision-making problem where the model actively chooses which parts of a video to attend to, storing discoveries in a running text memory rather than processing every frame. This approach decouples difficulty from video length — a 2-hour video is no harder than a 10-minute one if the agent identifies the right moments. The fact that performance improves with more reasoning turns (test-time scaling) suggests this agentic framing may generalize beyond video to other long-horizon tasks.

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

ReadSaveConnections

TxBench-PP tests whether AI agents can replicate the reasoning behind real drug-development decisions — things like interpreting assay data, ranking compounds, and planning next experiments — across 100 tasks drawn from 8 stages of preclinical pharmacology. The best system, Claude Opus, passed only 59% of tasks, meaning frontier models fail on roughly 4 in 10 realistic pharmaceutical reasoning steps. This puts a concrete number on the gap between 'can answer chemistry questions' and 'can actually assist drug discovery workflows.'

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

ReadSaveConnections

RNG-Bench uses two simple games — a card-matching task and a 3D maze — to test whether multimodal models can remember and use information from earlier in a sequence, a property called non-Markov reasoning. The benchmark's 'Memory Gap' metric cleanly separates errors caused by forgetting from errors caused by bad decisions, and finds that forgetting dominates: models see relevant information but fail to retain it across ~128K-token contexts. Fine-tuning a smaller open model on game rollouts improved performance without hurting general capability, suggesting targeted training data can partially close this gap.

██████████ 0.9 long-context Preprint

ReadSaveConnections

This paper shows that millions of hours of unlabeled first-person human videos — cooking, cleaning, assembling — can serve as training signal for robot control models, without ever labeling a single robot action. A specialized compression model learns to encode motion patterns from human videos into a shared 'action vocabulary' that robots can then be fine-tuned on with as few as 50 demonstrations. By separating what the model intends to do (from the language model) from what the robot's sensors currently observe (from a frozen visual encoder), the system also reduces cases where the model hallucinates physically impossible actions.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

When a vision-language model is further trained to control a robot, it forgets some of what it originally knew — particularly richer semantic and commonsense concepts, while simple perceptual ones survive. The paper introduces a lightweight evaluation protocol that rephrases standard knowledge benchmarks as physical object-placement actions so robot models can be tested without natural language output. A mechanistic finding — that knowledge signals peak in middle layers but weaken in upper layers after action fine-tuning — points to where future training strategies might preserve general intelligence while adding motor control.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

WorldLines constructs household simulation episodes where a robot must track what it did, what changed in the environment, and what goals remain — over long multi-step interactions — then answer questions or complete tasks based on that history. Existing benchmarks test memory retrieval from text, but WorldLines forces agents to translate past observations into physical actions in a world that has since changed, a much harder problem. The accompanying ObsMem framework explicitly separates event memory, object state tracking, and belief updating to give agents a structured way to reason about their own history.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

CAPRA uses a team of specialized LLM agents to evaluate student software architecture reports, achieving 88.8% agreement with human graders on a structured 8-criterion rubric. Its key anti-hallucination mechanism — Evidence Anchoring — requires every claim to match the source document using fuzzy string matching, so the system cannot invent evidence that isn't there. The moderate human inter-rater agreement (kappa 0.58) is a useful reality check: the AI is not beating humans, it is performing at a level comparable to human disagreement, which may be the realistic ceiling for this task.

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

ReadSaveConnections

ThinkDeception reframes deception detection — identifying lies from audio, video, and text — as a reasoning task rather than a classification task, forcing the model to explicitly explain which modalities conflict and why that conflict signals deception. Training uses a four-stage curriculum that gradually increases difficulty, combined with a reinforcement learning reward that scores both the reasoning process and the final answer. Making the model articulate its evidence chain is what makes outputs interpretable, which matters practically because deception detection in real-world settings (legal, clinical) requires explainable decisions.

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

ReadSaveConnections

This position paper argues that the web was built for humans reading pages, not for AI agents fetching structured information — and that the resulting friction (CAPTCHAs, bot-blocking, scraped paywalls) creates a hidden cost on all LLM-powered applications. The paper proposes concrete infrastructure changes: agent identity headers analogous to browser User-Agent strings, and dual-layer content architectures that serve human and agent audiences from the same URLs without deception. A less-discussed risk the paper highlights is 'epistemic recursion' — AI-generated content being consumed by other AI agents, progressively detaching the web's knowledge base from human-verified ground truth.

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

ReadSaveConnections

This prospective study deployed ChatGPT 4.1 in a live nuclear medicine department to answer real patient queries in parallel with human physicians and administrative staff, then scored both sets of responses on 15 quality dimensions using two independent rater groups. The study is notable for using actual patient queries rather than curated test sets, which typically flatter AI performance. Results showing unstable sensitivity-specificity trade-offs in LLM responses underscore why unsupervised clinical deployment remains premature even for administratively oriented queries.

██████████ 0.9 hallucination-grounding Peer-reviewed

Read