All digests
ResearchersENArtificial Intelligencedaily

[Artificial Intelligence] Daily digest — 266 papers, 0 strong connections (2026-04-15)

DeepScience — Artificial Intelligence
DeepScience
Artificial Intelligence · Daily Digest
April 15, 2026
266
Papers
10/10
Roadblocks Active
2
Connections
⚡ Signal of the Day
• Hallucination and grounding in multimodal models dominated today, with at least five independent papers attacking the problem from distinct angles — decoding interventions, structured document reasoning, physical adversarial attacks, rubric-based preference tuning, and RAG restructuring.
• A theoretical result ('The Verification Tax') establishes a hard statistical floor on how cheaply AI calibration can be audited: below a threshold of roughly one expected error per sample budget, miscalibration is provably undetectable — a direct constraint on compliance and safety monitoring regimes.
• Watch the intersection of agent tool-use and multimodal grounding: three papers (LMM-Searcher, See-Point-Refine, Don't Show Pixels) converge on the idea that long-horizon agents need richer intermediate representations to bridge visual perception and language reasoning, and results are materialising at scale.
📄 Top 10 Papers
Decoding by Perturbation: Mitigating MLLM Hallucinations via Dynamic Textual Perturbation
This paper diagnoses a specific mechanism behind multimodal hallucinations: language models are hypersensitive to how a question is phrased, allowing text priors to override actual visual evidence during answer generation. The proposed fix, Decoding by Perturbation (DeP), probes for these language biases at inference time by perturbing the input text and using the divergence to down-weight text-driven predictions. This matters because it requires no retraining and targets the root cause — language dominance during decoding — rather than patching outputs after the fact.
█████████ 0.9 hallucination-grounding Preprint
DocSeeker: Structured Visual Reasoning with Evidence Grounding for Long Document Understanding
DocSeeker identifies two concrete failure modes in multimodal LLMs on long documents: relevant evidence is buried in noise (low signal-to-noise ratio), and training datasets only supervise the final answer rather than the evidence-finding steps. The solution is a two-stage training pipeline — first teaching a 7B model to locate and reason over evidence via chain-of-thought distillation, then reinforcing evidence localization accuracy with a modified policy optimization reward (EviGRPO). The result outperforms both open- and closed-source models on five multi-page document benchmarks, and code is publicly available.
█████████ 0.9 hallucination-grounding Preprint
The Verification Tax: Fundamental Limits of AI Auditing in the Rare-Error Regime
This paper derives a mathematical lower bound on how many test samples are needed to reliably detect miscalibration in an LLM: the required sample count scales as the cube root of the inverse error rate, meaning a model that fails 1-in-10,000 times needs roughly 46× more evaluation data than one failing 1-in-100. Critically, having the model evaluate itself provides literally zero information about its calibration. This is not a software engineering problem that better tooling can fix — it is a fundamental statistical limit that constrains AI compliance audits, safety certifications, and deployment monitoring at scale.
█████████ 0.9 reasoning-reliability Preprint
Towards Long-horizon Agentic Multimodal Search
LMM-Searcher solves a practical bottleneck for multimodal agents: as search conversations grow to 100+ turns, storing all images in context becomes computationally prohibitive. The paper offloads images to an external file system and replaces them with lightweight text identifiers, loading visuals on-demand only when needed — cutting context overhead while preserving the ability to reason over images. A 30B-parameter model fine-tuned on 12K synthesised multi-hop trajectories matches or exceeds prior systems on four multimodal search benchmarks.
██████████ 0.8 multimodal-understanding Preprint
Don't Show Pixels, Show Cues: Unlocking Visual Tool Reasoning in Language Models via Perception Programs
Even when vision tools (depth estimators, optical flow detectors) produce correct outputs, multimodal LLMs often fail to use that information — because raw pixel-level tool outputs are poorly matched to how language models process information. Perception Programs (P²) converts these dense outputs into compact, structured text summaries with normalised spatial coordinates, requiring no model retraining. Tested across six perception tasks, P² delivers a 22% average accuracy gain with GPT-4o Mini, and the code is publicly released.
██████████ 0.8 agent-tool-use Preprint
Challenging Vision-Language Models with Physically Deployable Multimodal Semantic Lighting Attacks
This paper demonstrates that adversarial lighting patterns — triangular projections optimised in a 9-dimensional parameter space using genetic algorithms — can reliably fool production vision-language models like CLIP, LLaVA, and BLIP in the real physical world, not just in digital simulations. The attacks transfer across model architectures and trigger severe semantic hallucinations in captioning and visual question-answering tasks. This is directly relevant for any deployment of vision-enabled agents in physical environments, as it establishes a low-cost attack vector that bypasses standard digital adversarial defences.
██████████ 0.8 hallucination-grounding Preprint
Visual Preference Optimization with Rubric Rewards
Rather than scoring model outputs as simply correct or incorrect, this paper creates instance-specific rubrics — per-image checklists of what a good answer must include — and uses these as reward signals for preference learning (rDPO). A 30B open-source judge scoring against rubrics comes close to GPT-4-level reward modelling, and the resulting model scores 61.01 on a comprehensive multimodal benchmark versus 52.36 for a style-constrained baseline. The key insight is that outcome-based filtering alone actually degrades performance (75.82) compared to rubric-based filtering (82.69), suggesting that answer quality requires richer evaluation criteria than binary correctness.
██████████ 0.8 multimodal-understanding Preprint
See, Point, Refine: Multi-Turn Approach to GUI Grounding with Visual Feedback
GUI grounding — having an AI agent click on the correct pixel in a user interface — typically fails because single-shot predictions cannot recover from small spatial errors in dense UIs like code editors. This paper replaces single-shot prediction with an iterative closed-loop process: the agent points, observes where it actually landed, and refines its click based on visual feedback. Multi-turn refinement significantly outperforms state-of-the-art single-shot models on click precision and task success, including in dynamic UIs where layout changes between turns.
██████████ 0.8 agent-tool-use Preprint
ARGOS: Who, Where, and When in Agentic Multi-Camera Person Search
ARGOS reframes multi-camera person search — tracking individuals across surveillance feeds — as a reasoning problem with information asymmetry: the agent must ask limited questions of a witness to narrow down identity, location, and time of sighting. The benchmark contains 2,691 tasks across 14 real-world scenarios encoded in a Spatio-Temporal Topology Graph. Even the best model (GPT-4o family) achieves only 0.383 Turn-Weighted Success on the harder tracks, confirming the benchmark is far from solved and highlighting a gap between current LLM reasoning and real investigative workflows.
██████████ 0.8 agent-tool-use Preprint
IDEA: An Interpretable and Editable Decision-Making Framework for LLMs via Verbal-to-Numeric Calibration
IDEA extracts an LLM's decision-making knowledge into an explicit parametric model — mapping verbal expressions like 'likely' or 'rarely' to calibrated probabilities — by jointly learning those mappings and decision weights via an Expectation-Maximisation algorithm. This makes the model's reasoning auditable and editable without retraining: a user can inspect or override specific factor weights. A Qwen-3-32B backbone with IDEA scores 78.6% on decision benchmarks, outperforming both DeepSeek R1 (68.1%) and GPT-4-class models (77.9%), with code publicly released.
██████████ 0.8 interpretability Preprint
🔬 Roadblock Activity
Roadblock Papers Status Signal
Efficiency & Scaling 110 Active Largest paper pool today; SNN-Synthesis v8 offered a notable empirical result that 4-bit quantization noise can act as stochastic resonance, with a 1.5B quantized model beating a full-precision 7B baseline on ARC-AGI-3.
Reasoning Reliability 105 Active A theoretical ceiling on calibration auditing ('Verification Tax') sets a hard constraint on how reliably reasoning errors can be detected at scale, directly challenging assumptions behind current LLM evaluation pipelines.
Multimodal Understanding 93 Active Multiple convergent approaches — file-based visual offloading (LMM-Searcher), structured perception cues (P²), rubric-based preference learning (rDPO), and long-document evidence grounding (DocSeeker) — all address the same bottleneck: language models cannot natively process raw visual representations at scale.
Hallucination & Grounding 74 Active DeP (decoding-time perturbation) and MSLA (physical adversarial lighting) together illustrate that hallucination is attackable from both the output side (decoding intervention) and the input side (adversarial perception), with neither requiring model retraining.
Agent Tool Use 61 Active Three papers (See-Point-Refine, LMM-Searcher, P²) and two scored connections converge on the same failure mode: agents cannot reliably translate visual perception into correct tool invocations without intermediate structured representations or iterative feedback loops.
Interpretability 53 Active IDEA's verbal-to-numeric calibration framework offers a practical path to auditable LLM decision-making without requiring retraining, achieving competitive benchmark performance while making factor weights human-readable and editable.
Data Quality & Curation 44 Active Moderate activity with no standout paper in the top 10; DocSeeker's ALR chain-of-thought distillation and ARGOS's deterministic ground-truth generation pipeline are the closest adjacent contributions.
Alignment & Safety 40 Active RePAIR's interactive machine unlearning framework proposes inference-time knowledge erasure via closed-form MLP activation redirection, though low reproducibility limits near-term practical uptake.
Embodied AI 26 Active VULCAN applies vision-language models to multi-agent fire disaster navigation, highlighting that standard indoor navigation benchmarks do not capture the failure modes introduced by smoke, heat, and dynamic environments.
Long Context 17 Active Quieter day for long-context work specifically; DocSeeker and LMM-Searcher both address related scaling problems (multi-page documents, 100-turn agent trajectories) but are categorised under their primary roadblocks.
View Full Analysis
DeepScience — Cross-domain scientific intelligence
Sources: arXiv · OpenAlex · Unpaywall
deepsci.io