290

Papers

10/10

Roadblocks Active

3

Connections

VeriGraph reformulates LLM agent reasoning as the construction of an explicit directed acyclic graph of evidence, where every claim must attach to a computational, grounding, or derivational node before it can propagate. An 8B model trained with graph-based reinforcement learning achieves the highest overall score on four data-analytics benchmarks and a grounding rate of 87.6%, compared to baselines that generate plausible-sounding but unanchored conclusions. This matters because it replaces trust in fluent-sounding output with a checkable artifact — auditors and downstream systems can inspect the graph rather than re-run inference.

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

ReadSaveConnections

Vision-language models confidently answer questions even when the visual evidence they need is absent or contradicted — a systematic overconfidence problem in embodied settings. Semantic Flip fixes this by synthesizing training pairs where either the question or the video memory is deliberately corrupted, then training a lightweight rejection gate on top of a frozen model without touching the underlying weights. The approach generalises to new models cheaply and publicly releases code, making it a practical plug-in for any deployment that needs the model to say 'I don't know' reliably.

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

ReadSaveConnections

Stronger code-reasoning models reliably exhibit specific token-level signals — verification steps, backtracking, backward chaining — that weaker models lack, and adding these patterns at inference time improves performance on math, ordering, and optimisation tasks. The paper shows these same cognitive behaviours also improve both supervised fine-tuning and reinforcement learning when injected during training. This is practically useful because it gives a cheap diagnostic for whether a model will reason reliably with code tools, and a concrete intervention that doesn't require more data.

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

ReadSaveConnections

Kairos proposes treating world models as first-class operational infrastructure for robots rather than as auxiliary visual generators, combining a three-stage cross-embodiment training curriculum with a hybrid attention mechanism designed to maintain coherent state across long action sequences. The architecture provides formal error-accumulation bounds and releases model weights publicly, which matters because most prior embodied world models are evaluated only in narrow domains and not released. If the benchmark claims hold up, this is a credible open alternative to proprietary physical-AI stacks.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

Qwen-RobotWorld trains a 60-layer diffusion transformer on 8.6 million video-text pairs spanning manipulation, driving, navigation, and human-to-robot transfer, using a frozen Qwen2.5-VL encoder to process language action commands. It ranks first on EWMBench and DreamGen Bench and outperforms all open-source models on two additional benchmarks, demonstrating that a unified language-conditioned generation approach can handle qualitatively different physical domains in a single model. The dataset and training pipeline are proprietary, which is a significant reproducibility barrier, but the architecture description is detailed enough to guide independent reimplementation.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

A small binary classifier trained to judge whether a reasoning trace is correct can, via entropy-based confidence thresholding, generate reliable pseudo-labels for large pools of unlabeled problems — achieving accuracy comparable to using 10–15 times as many human-labeled examples on math and visual reasoning tasks. The key mechanism is that the classifier's own uncertainty (measured by output entropy) acts as a quality filter, keeping only high-confidence pseudo-labels for fine-tuning the generator. This matters for practitioners who cannot afford large labeled datasets: it dramatically lowers the annotation cost of improving reasoning quality.

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

ReadSaveConnections

Standard benchmarks and single-model judges miss systematic agent failures like phantom tool-call claims, silent policy drift, and refusals that look safe but never resolve the task. Human-on-the-Bridge runs smaller 'Harness' LLMs as adversarial evaluators against frontier agents over multi-turn conversations, with multi-juror scoring and evidence-linked failure reports — finding failures invisible to static benchmarks across 23,500 agent turns in finance, healthcare, and code domains. Code is publicly released, making it a practical infrastructure for teams that need to stress-test agents before deployment rather than after.

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

ReadSaveConnections

LabOSBench provides 96 browser-based tasks across eight simulated scientific instruments — electron microscopes, X-ray diffractometers, atom probes — and finds that current vision-language models handle isolated GUI subtasks but collapse on feedback-driven, multi-step workflows that require interpreting instrument readings to decide the next action. The web-based design avoids OS virtualisation, making it more reproducible than prior GUI benchmarks, and the results expose a concrete gap between agentic AI capability and real laboratory automation needs. This is the first benchmark specifically targeting scientific instrument GUIs, which are a realistic near-term deployment target.

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

ReadSaveConnections

MyPCBench tests computer-use agents on tasks drawn from real user requests, requiring personalisation across multiple applications in a simulated Linux desktop environment. The best model, Claude Opus 4.6, completes only 55.4% of tasks — the only model above 50% — with failures concentrated on multi-application workflows where personalised context must be maintained across a long action trajectory. The benchmark quantifies the gap between current impersonal evaluation setups and what personal assistant deployment actually demands, providing a concrete target for long-context and memory research.

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

ReadSaveConnections

LLMs frequently fail to identify the single decisive piece of evidence in a long context — a tool-call log line, a subtle image detail — even when the overall reasoning looks plausible. ContextRL adds a contrastive auxiliary training objective to standard GRPO post-training: the model must prefer the context that actually supports a correct answer over a minimally perturbed confounding alternative, achieving +2.2% on five long-horizon agentic benchmarks and +1.8% across twelve visual QA datasets. The mechanism is lightweight relative to its gains and applies to both text and image contexts, though reproducibility depends on GPT-5.4 for constructing training pairs.

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

ReadSaveConnections