Who Knows What - and What Does It Support? Perspective-Bounded Evidence Calibration for Multimodal Theory of Mind
Calibrates multimodal theory-of-mind inference to the evidence each observer can actually see.
AI Storyline
15 items · 6 sources · 9 days
Operational story trace
Reasoning Multimodal
Follow in this browser to see new updates on your Live feed.
Latest change
A Show HN project, Receipts, ships agent skills that make PRDs and metrics cite their evidence, a tooling counterpart to the thread's papers on agents that must back answers with support.
Earlier contextThe story so far
A loose research cluster, not one event: from Oct 23, arXiv papers on multimodal models gathered around two problems, weak visual perception and answers that lack supporting evidence. Early November added hallucination mitigation, unlearning and knowledge editing, then a human-perception benchmark and an evidence-composing memory system for multimodal agents.
editor-curated · source-linked
Arc
Oct 27Nov 12 · now
EVIDENCE · Oct 27–30Correct is not the same as supported4 sources · show sources ▾
Clearing the FoG: Training KG Agents to Retrieve the Evidence Behind Their Answers
Trains knowledge-graph agents to retrieve the evidence behind an answer, since a correct answer may be unsupported.
When Visual Evidence Is Insufficient: How Instructions, Post-Training, and Internal Activations Shape Refusal in Multimodal Language Models
Studies how instructions, post-training and activations shape refusal when an image gives too little evidence.
When AI Investigators Cannot Trust Their Tools: Conflict-Aware Evidence Verification for Cyber Threat Investigation
Security angle: alert-triage agents that assume their detection tools are trustworthy get a conflict-aware verification step.
PERCEPTION · Oct 23 – Nov 9Visual perception named as the bottleneck for multimodal reasoning6 sources · show sources ▾
Beyond Post-Hoc Rewards: Privileged Reasoning Exploration for Multimodal Embeddings
Opens the cluster: explores reasoning during training of multimodal embeddings instead of rewarding it after the fact.
PerceptRL: Making Visual Perception Rewardable in MLLMs from Answers to Interactions
PerceptRL makes fine-grained visual perception a rewardable target in MLLM post-training.
Latent-Lens: Visual Perception in Small Language Models Through Latent Communication
Replaces lossy text hand-offs between a vision model and a small language model with latent communication.
Latent Reasoning in Continuous Space for Unified Multimodal Models
Moves chain-of-thought for unified multimodal models into continuous latent space.
MIRA: Reinforcing Multimodal Reasoning via Deceptive Contextual Augmentation
MIRA targets the perception-reasoning gap in RLVR training with deceptive contextual augmentation.
HumanSightBench: A Benchmark for Human Visual Perception in MLLMs
HumanSightBench benchmarks MLLMs on visual questions people solve easily.
ADJACENT · Nov 3–6Loosely linked: hallucination, unlearning and editing inside multimodal models3 sources · show sources ▾
VIGOR: Visual Gain Ordering for Hallucination Mitigation in Multimodal Discrete Diffusion Language Models
Side paper: hallucination mitigation for multimodal discrete diffusion language models.
Reasoning-Aware Cross-Modal Association Unlearning in Multimodal Large Reasoning Models
Side paper on unlearning cross-modal associations in multimodal reasoning models.
ROSE: Risk-Aware Orthogonal Subspace Navigation for Lifelong Knowledge Editing in Multimodal Large Language Models
Side paper on lifelong knowledge editing in multimodal LLMs.
NOW · Nov 9–12Evidence handling reaches agent memory and agent skills2 sources · show sources ▾
HiVe-Mem: Interleaving Horizontal Memory Expansion and Vertical Evidence Composition for Multimodal Agents
HiVe-Mem pairs memory expansion with evidence composition for multimodal agents with long interaction histories.
Show HN: Receipts, agent skills that make PRDs and metrics cite their evidence
First tooling item: agent skills that make PRDs and metrics cite their evidence.
Day 1 Friday, Oct 23, 2026
Beyond Post-Hoc Rewards: Privileged Reasoning Exploration for Multimodal Embeddings
Opens the cluster: explores reasoning during training of multimodal embeddings instead of rewarding it after the fact.
Day 2 Tuesday, Oct 27, 2026
Who Knows What - and What Does It Support? Perspective-Bounded Evidence Calibration for Multimodal Theory of Mind
Calibrates multimodal theory-of-mind inference to the evidence each observer can actually see.
Clearing the FoG: Training KG Agents to Retrieve the Evidence Behind Their Answers
Trains knowledge-graph agents to retrieve the evidence behind an answer, since a correct answer may be unsupported.
When Visual Evidence Is Insufficient: How Instructions, Post-Training, and Internal Activations Shape Refusal in Multimodal Language Models
Studies how instructions, post-training and activations shape refusal when an image gives too little evidence.
PerceptRL: Making Visual Perception Rewardable in MLLMs from Answers to Interactions
PerceptRL makes fine-grained visual perception a rewardable target in MLLM post-training.
Day 3 Friday, Oct 30, 2026
When AI Investigators Cannot Trust Their Tools: Conflict-Aware Evidence Verification for Cyber Threat Investigation
Security angle: alert-triage agents that assume their detection tools are trustworthy get a conflict-aware verification step.
Latent-Lens: Visual Perception in Small Language Models Through Latent Communication
Replaces lossy text hand-offs between a vision model and a small language model with latent communication.
Latent Reasoning in Continuous Space for Unified Multimodal Models
Moves chain-of-thought for unified multimodal models into continuous latent space.
Day 4 Monday, Nov 2, 2026
MIRA: Reinforcing Multimodal Reasoning via Deceptive Contextual Augmentation
MIRA targets the perception-reasoning gap in RLVR training with deceptive contextual augmentation.
Day 5 Tuesday, Nov 3, 2026
VIGOR: Visual Gain Ordering for Hallucination Mitigation in Multimodal Discrete Diffusion Language Models
Side paper: hallucination mitigation for multimodal discrete diffusion language models.
Day 6 Wednesday, Nov 4, 2026
Reasoning-Aware Cross-Modal Association Unlearning in Multimodal Large Reasoning Models
Side paper on unlearning cross-modal associations in multimodal reasoning models.
Day 7 Friday, Nov 6, 2026
ROSE: Risk-Aware Orthogonal Subspace Navigation for Lifelong Knowledge Editing in Multimodal Large Language Models
Side paper on lifelong knowledge editing in multimodal LLMs.
Day 8 Monday, Nov 9, 2026
HumanSightBench: A Benchmark for Human Visual Perception in MLLMs
HumanSightBench benchmarks MLLMs on visual questions people solve easily.
HiVe-Mem: Interleaving Horizontal Memory Expansion and Vertical Evidence Composition for Multimodal Agents
HiVe-Mem pairs memory expansion with evidence composition for multimodal agents with long interaction histories.
Day 9 Thursday, Nov 12, 2026
Show HN: Receipts, agent skills that make PRDs and metrics cite their evidence
First tooling item: agent skills that make PRDs and metrics cite their evidence.
What to watch — open questions
- Do KG agents trained to retrieve supporting evidence keep their answer accuracy on standard QA benchmarks?
- Does conflict-aware evidence verification get tested on production security alert triage, beyond the paper's own evaluation?
- How do frontier MLLMs score on HumanSightBench compared with human performance?
- Do refusal behaviours for insufficient visual evidence survive post-training, or does fine-tuning erode them?
- Does Receipts publish any measurement of how often its cited evidence actually supports the claim?
How this thread was built
editor wrote the arc · 4 beats
Storylines are threaded mechanically from the feed: stories that share a distinctive anchor across multiple days and sources. Each item links to its original source. The evidence trace, current state, and open questions are written by the editor routine and refreshed whenever a new beat lands.