288

Papers

11/11

Roadblocks Active

4

Connections

Five leading vision-language models — GPT-5, Gemini 2.5 Pro, o3, GLM-4.5V, and Qwen 2.5 VL — were systematically tested on medical image question-answering tasks requiring both visual localization and clinical reasoning. Even the best model could only correctly identify anatomical regions 19.1% of the time at standard overlap thresholds, and all models showed clinically dangerous confusion between left and right anatomy. The result matters because it exposes that current VLMs treat visual grounding and language reasoning as loosely coupled modules — a design that works for general tasks but fails when precise spatial evidence is required for a safe clinical decision.

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

ReadSaveConnections

ANCORA trains a single model to alternate between generating verifiable problems and solving them, without any human-labeled data, by using a two-level reward mechanism that links the quality of the question to the quality of the answer. A key practical fix is an iterative self-distillation step that constrains the model to a valid output space before reinforcement learning begins, preventing the training from collapsing under sparse verification signals. This matters because it opens a path to continuously improving reasoning capabilities using only model self-interaction — reducing dependence on expensive human-curated problem sets.

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

ReadSaveConnections

This paper injects controlled misinformation into the shared information artifacts that agents exchange, then tracks how contamination propagates through execution traces across three different language models. A surprising finding is that workflow traces can look nearly identical to clean runs while producing wrong outputs, and vice versa — meaning structural similarity to a correct trace is not a reliable signal of correctness. Current verification guardrails tested in the study failed to catch contamination, which is a direct concern for any multi-agent deployment where agents read and write shared state.

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

ReadSaveConnections

PRISM identifies a concrete problem in the standard supervised fine-tuning → reinforcement learning pipeline for multimodal models: SFT shifts the model's output distribution away from both the original model and the supervision target, creating drift that compounds when RL begins. The fix is a distribution-alignment stage inserted between SFT and RL, formulated as an adversarial game between the policy and a mixture-of-experts discriminator with separate perception and reasoning heads, using only on-policy rollouts with no access to teacher logits. Tested on Qwen3-VL at two scales across three RL algorithms, PRISM consistently improves downstream performance — suggesting the SFT-RL gap is a general structural problem, not a model-specific artifact.

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

ReadSaveConnections

FlashRT makes optimization-based adversarial attacks on long-context language models practical by combining selective token recomputation and partial-context gradient approximation, reducing GPU memory from 264 GB to 66 GB and running 2–7× faster than the prior baseline. The key insight is that most of the computational cost in iterative attack optimization is redundant — the majority of the context does not need full gradient recalculation at every step. Beyond security, the paper matters because the same efficiency techniques could accelerate any iterative optimization loop over long sequences, including chain-of-thought refinement and test-time compute methods.

██████████ 0.8 long-context Preprint

ReadSaveConnections

The paper generates 1,000 realistic synthetic computer environments — complete with folder hierarchies, files, and persona-specific content — and runs simulated month-scale work sessions of 2,000+ turns per environment to produce training data for productivity agents. Agents trained on these simulations improve on both in-domain and out-of-domain benchmarks, suggesting that realistic environmental richness matters more for agent learning than simply increasing the number of short interactions. This addresses a key bottleneck in agent training: the scarcity of long-horizon, contextually grounded interaction data at scale.

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

ReadSaveConnections

Echo-α builds a medical AI agent that calls organ-specific object detectors as tools during ultrasound interpretation, combining their localized outputs with global visual context inside a reasoning loop rather than relying on a single monolithic model. Training uses a nine-task curriculum followed by reinforcement learning with separate reward configurations that produce two specialized variants — one optimized for lesion localization (56.7% F1 at standard overlap) and one for diagnosis (74.9% accuracy on cross-center renal data). The architecture demonstrates that decomposing perception into callable specialist tools improves both transparency and performance in domains where spatial precision is medically critical.

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

ReadSaveConnections

This survey maps how reinforcement learning for LLMs is shifting from narrow task-specific agents with hand-designed rewards toward agents that set their own goals, plan over extended horizons, reflect on their behavior, and adapt strategies dynamically. The paper distinguishes agentic RL from classical RL by the inclusion of cognitive-like mechanisms — meta-reasoning, self-reflection, and multi-step decision-making — inside the learning loop itself. For practitioners, it provides a structured taxonomy of current approaches and open challenges, which is useful for identifying where the field's coverage is thin versus where it is converging.

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

ReadSaveConnections

This survey examines why supervised fine-tuning alone breaks down for agents navigating graphical user interfaces, focusing on three failure modes: inability to assign credit across long action sequences, brittleness when the interface distribution shifts, and the danger of irreversible actions during exploration. The survey argues that composite, multi-tier reward architectures are emerging as the practical solution, and that latency bottlenecks in real-environment interaction are pushing the field toward world-model-based training where agents simulate GUI states rather than executing them. The framing is useful for understanding why GUI agents are harder than benchmark performance suggests.

██████████ 0.7 agent-tool-use Preprint

ReadSaveConnections

The paper proposes a pipeline where an LLM decomposes high-level natural-language safety goals into candidate causal rules expressed in first-order logic, then a separate verification engine checks those rules for syntax errors, schema violations, logical consistency, and safety constraint satisfaction before they are integrated into an autonomous driving system. The approach is evaluated on two driving scenarios and successfully derives rule sets from goals without manual specification. The practical value is in the verification stage: it provides a structured rejection mechanism that catches rule failures before deployment rather than after, which is a meaningful step toward auditable AI behavior in safety-critical domains.

██████████ 0.7 alignment-safety Preprint

ReadSaveConnections