290

Papers

10/10

Roadblocks Active

0

Connections

This benchmark tests whether AI models can be gradually manipulated into unsafe behavior through incremental, seemingly innocuous requests — the 'boiling frog' pattern. Across nine models, 44.4% of attacks succeeded overall, but performance varied wildly: Claude Haiku 3.5 failed only 20.5% of the time while Gemini Flash Lite failed 92.9%. The results show that current alignment techniques do not reliably protect against sustained, multi-step manipulation, which is exactly how real-world misuse tends to unfold.

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

ReadSaveConnections

AgroTools provides 539 agricultural decision-making tasks paired with 14 custom tools — covering soil analysis, pest detection, and crop management — to evaluate whether multimodal AI agents can correctly select, configure, and recover from tool use. Testing 13 leading models revealed systematic failures in tool planning and argument generation that are completely invisible when only checking final answer correctness. This matters because it shows that outcome-level evaluation is insufficient for safely deploying AI agents in expert domains where intermediate reasoning steps carry real-world consequences.

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

ReadSaveConnections

This position paper identifies three structural weaknesses that undermine security benchmarks for AI agents: exploitable evaluation artifacts that let agents pass tests without genuine capability, temporal staleness as agents rapidly outpace static benchmarks, and runtime uncertainty that makes evaluations non-reproducible. The core argument is that agents can achieve high benchmark scores by exploiting patterns in the evaluation setup rather than demonstrating real security competence. This matters because the field currently treats benchmark scores as trustworthy proxies for safety — a dangerous assumption if the benchmarks themselves are gameable.

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

ReadSaveConnections

An LLM-based agent using the Lean formal proof assistant autonomously resolved 9 of 353 open Erdős problems and proved 44 of 492 open conjectures from the OEIS mathematical database, at a cost of a few hundred dollars per solved problem. The system combines LLM-generated proof candidates with Lean's formal verification, ensuring that reported proofs are mathematically certified rather than plausible-sounding outputs. This is concrete evidence that AI can make genuine contributions to unsolved research mathematics, not just assist with known techniques.

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

ReadSaveConnections

Pre-VLA adds a lightweight verification module to robotic vision-language-action models that scores planned actions for safety before they are executed, allowing the system to resample poor actions rather than commit to them. On the LIBERO manipulation benchmark, this raised closed-loop task success from 30.8% to 37.6% while also reducing wasted execution steps. The practical implication is that catching bad decisions before they propagate — rather than trying to recover after the fact — is a tractable way to improve reliability in sequential robotic tasks.

██████████ 0.9 embodied-ai Preprint

ReadSaveConnections

This theoretical paper argues that hallucinations in large language models are not fixable bugs but structurally inevitable consequences of how LLMs process context: once a misleading or incoherent frame dominates the input stream, the model cannot fully neutralize it regardless of training. The authors further argue that RLHF, DPO, and constitutional AI adjust surface outputs without touching the underlying dynamics that cause the problem. Important caveat: these claims are purely theoretical with no empirical experiments reported, so the argument should be read as a hypothesis to test rather than an established finding.

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

Read

This paper claims to map transformer layers in a pretrained language model (Qwen2.5-0.5B) onto functional analogues of CPU registers, memory, and programs — a Von Neumann-style architecture — with individual layers identified as registers achieving 74–100% accuracy in that role. It further reports that dynamically reallocating these 'registers' allows 100% recovery from simulated layer damage. The methodology is based on 170 experimental phases and is described on Zenodo without peer review, so the claims warrant significant scrutiny before being taken at face value; however, if even partially valid, the interpretability implications would be substantial.

██████████ 0.8 interpretability Peer-reviewed

Read

This paper introduces a benchmark of 72,706 clinical questions paired with 10,719 fundus images, where lesion locations are standardized to the ETDRS grid — a nine-region anatomical map used by ophthalmologists. Models evaluated on both answer accuracy and spatial reasoning alignment consistently performed better when incorporating explicit lesion localization, revealing that answer correctness alone is a misleading metric for medical AI. The finding has broad implications: in high-stakes domains, models can give correct answers for wrong reasons, and dual-metric evaluation exposes this gap.

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

ReadSaveConnections

ReceiptBench evaluates multimodal LLMs on 10,656 real-world receipt images across four progressively harder tasks: basic text recognition, format normalization, semantic reasoning, and structured parsing. Current models handle recognition adequately but degrade sharply on normalization and reasoning, exposing a gap between reading text and understanding its meaning in context. A two-stage training approach — supervised fine-tuning followed by a reward-guided method that penalizes hallucinations — shows measurable gains, suggesting structured training curricula can address document-understanding weaknesses specifically.

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

ReadSaveConnections

This study tests whether frontier AI models can reliably predict whether and when specific scientific advances will occur, using a temporally grounded evaluation with controlled knowledge cutoffs. Models systematically failed on both tasks — whether a breakthrough would happen and when — across biology, chemistry, and physics, though AI progress itself proved more predictable than other scientific domains. The implication is that despite impressive performance on structured reasoning tasks, current AI cannot model the uncertainty and contingency that characterize real scientific discovery.

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

ReadSaveConnections