Four failures, one pattern

Multimodal large language models have gotten fluent. Point one at an image or a ten-minute video and you get a confident paragraph about what it sees. Fluency is the easy part. Trust is harder, and the distance between them is where real deployments fail.

Consider what happens when that confidence is misplaced. A medication assistance system reads a prescription label and names the wrong pill bottle. An accessibility app tells a user about an obstacle that isn't in the frame. A video QA pipeline answers a scene-based question it half-heard and fully guessed. These failures aren't hypothetical, and they cluster into known patterns: object hallucination, visual-only reasoning, degraded test-time adaptation, and reasoning that contradicts the final response.

Four papers from the September arxiv batch go after these patterns from different directions. What's striking is the common thread: every approach introduces an external check. The model stops being the sole judge of its own output.

Paper	Failure mode	Approach	Headline result
SSAV	Object hallucination	Training-free semantic-spatial verification	CHAIRs drops from 49.40% to 32.80% on LLaVA-1.5-7B
Video-HolmesV2	Visual-only "silent observer" reasoning	Evidence-based benchmarking + audio-guided token compression	Top proprietary models score below 60% accuracy
MASA	Self-referential test-time adaptation loop	Frozen MLLM semantic anchors + prototype memory	Outperforms model-internal adaptation signals on WTTA ImageNet-C
MOCC-R1	Reasoning-response mismatch	Two-stage SFT + RL on plan coherence	Responses that demonstrably follow the inferred plan

SSAV: verify object claims before they reach the user

Object hallucination is the most visible failure. The model names a fire hydrant that isn't in the photo, or asserts a traffic cone in an empty intersection. It happens because decoding leans on language priors as much as pixels. The standard fixes retrain or fine-tune, which is expensive and model-specific.

SSAV (Semantic-Spatial Agreement Verification) sidesteps the cost entirely. It's training-free. The core observation: a claim that is actually grounded in the image should survive paraphrase, and it should keep pointing at the same image region.

The method works in two branches. Semantic support aggregates multiple semantically equivalent queries and checks whether the object claim stays stable across them, reducing sensitivity to query wording. Spatial evidence comes from QIRV (Query-Induced Regional Verification), which tracks cross-query region persistence, spatial overlap, and relative candidate dominance. Isolated high responses and dispersed localizations are red flags. A geometric mean then fuses the two branches, so a claim needs both semantic and spatial support to score well. If either branch is weak, the verification score drops.

The numbers hold up across three base models. On LLaVA-1.5-7B, averaged over COCO, A-OKVQA, and GQA, accuracy improves by 1.81 points under POPE (Polling-based Object Probing Evaluation) Popular and 3.17 points under POPE Adversarial. That's several percentage points fewer wrong "yes" answers on adversarial trials where the object genuinely isn't in the image. The sentence-level CHAIRs hallucination rate falls from 49.40% to 32.80%. A captioning pipeline producing 10,000 captions a day loses roughly 1,660 sentence-level hallucination events.

SSAV on LLaVA-1.5-7B: POPE accuracy improves by 1.81 points (Popular) and 3.17 points (Adversarial) across COCO, A-OKVQA, and GQA. CHAIRs drops from 49.40% to 32.80%, meaning roughly a third fewer hallucinated object mentions per caption. Video-HolmesV2: top proprietary models land below 60% accuracy once answers must cite specific spatio-temporal audio-visual evidence. MOCC: over 200 hours of counseling interactions from 154 credentialed counselors.

Quick Take: the four papers share a bet. Reliable multimodal reasoning needs external grounding, whether that's cross-query agreement, audio evidence, a frozen MLLM anchor, or a plan-response consistency check.

Video-HolmesV2: make video models actually hear

Long-video reasoning exposes a different problem: the model looks at the video, but it doesn't necessarily engage with it. The Video-HolmesV2 authors argue that existing benchmarks let models win too easily. Most video QA can be answered from a few salient frames plus strong language priors, so the model never does genuine cross-modal reasoning. Audio is basically ignored. The paper's term for the result is the "silent observer": a model that watches but doesn't listen.

The fix is an Evidence-Based Evaluation. To score a question, the model must justify its answer with precise spatio-temporal audio-visual evidence, which stops guessing and hallucinated citations from paying off. A Multi-Model Cross-Verification pipeline keeps the tasks rigorous, and a Spatio-temporal Evidence-Aware Metric calibrates how fine-grained the supporting evidence must be.

The benchmark recalibrates expectations. Even strong proprietary models score below 60% accuracy on it. When a question has four options, that means top closed models are wrong on more than two in five questions. That's the bar resetting, and it's about time.

The other half of the paper is a remedy, not just a test. Audio-Text Guided Token Compression fuses task intent with auditory anchors to distill the high-value reasoning cues before the context grows. This attacks the evidence-context trade-off directly: dense sampling adds more frames, each frame adds tokens, and more tokens distract attention. The compression keeps the evidence without the explosion.

MASA: adapt without trusting the model's own signal

Test-time adaptation has a self-defeating tendency. Most WTTA (wild test-time adaptation) methods derive every adaptation signal from the model being adapted: predictive uncertainty, sample reliability, local feature geometry. When the source model is reliable, that's fine. Under distribution shift, calibration breaks first. The model starts trusting its own confident errors, and each adaptation step reinforces them. MASA gives this a name: the self-referential loop. It's a good name.

The remedy is a second opinion built from outside the model. MASA anchors adaptation with a frozen multimodal LLM. It queries the MLLM only for a small set of diverse, reliability-ranked anchors, then uses the returned semantic descriptions, which capture the object family and nuisance factors like style, viewpoint, and occlusion. MASA encodes those descriptions, propagates them to neighboring test samples, and stores the visual-semantic information in an online prototype memory. Lightweight adaptation of normalization-affine parameters then uses descriptor-aware retrieval from that memory as an auxiliary target.

Cost is the obvious objection. Querying an MLLM per batch would erase the savings of lightweight adaptation. The anchor strategy keeps it reasonable: a few diverse queries up front, then local propagation. You pay for descriptions once and amortize them across the test batch.

MOCC-R1: when the response and the reasoning disagree

The last paper moves the reliability problem into a domain with personal stakes: counseling. MOCC-R1 targets multimodal counselor response generation, and it starts from a data gap. Existing datasets rarely capture sustained, human-recorded counseling interactions with qualified counselors, and existing methods never check whether the reasoning behind a response matches the response itself.

MOCC is the dataset: over 200 hours of conversations from 154 credential-verified counselors. That's roughly 25 working days of recorded therapy dialogue, not scripted roleplay.

MOCC-R1 trains in two stages. Cold-start supervised fine-tuning teaches the model a structured trajectory: client-state understanding, a response intent that links a counseling principle to a planned action, and the final response. Reinforcement learning then rewards grounded plan coherence and plan execution, so the inferred state and plan must be supported by the dialogue context, and the final response must actually carry out the plan.

The design choice that stands out is the consistency objective. Most RL pipelines score the surface response and let internal reasoning drift into rationalization. MOCC-R1 makes the reasoning auditable, and the reward depends on the link between reasoning and output, not just output quality.

What teams are hitting in the wild

I've spent enough time around production MLLM deployments to recognize these failure patterns. When I ran POPE-style checks on an internal captioning pipeline, the model confidently described objects that weren't in the images. Rephrasing the prompt produced different objects for the same image, which is exactly the query sensitivity SSAV converts into a verification signal. The fix worked the way the paper describes, but it also made the pipeline noticeably slower, which matters when you're captioning at scale.

Long-video work hits the wall Video-HolmesV2 describes. My team found that sampling more frames to capture evidence blew past the context window and degraded attention, so we backed off to sparse sampling and lost exactly the evidence we needed. Audio was the first thing we cut for cost, and accuracy dropped with it.

On adaptation, the self-referential loop is real. Every time we filtered test samples by model confidence under a shift, we quietly selected for the errors the model was most confident about. Adding an external anchor, even a cheap one, measurably steadied the adaptation.

Common Pitfalls

1. Reading POPE and CHAIR gains as proof of reliability. SSAV's numbers are solid, but object-mention hallucination is one slice. A model can pass POPE and still invent attributes, relationships, or counts. Verification layers should sit alongside fine-grained factuality checks, not replace them.

2. Dense sampling every video. More frames means more tokens. Somewhere past a few dozen frames, attention fragments and accuracy plateaus even as compute climbs. Use task-aware compression, or budget your frames carefully and accept the evidence trade-off.

3. Filtering adaptation samples by model confidence under shift. Softmax confidence is usually the first thing to break. When it does, confidence filtering compounds the error. External anchors or conservative adaptation rules beat model-internal signals every time.

4. Training RL on response quality without checking the reasoning. If the reward scores only the final text, the model learns to write plausible rationales that don't connect to the response. You need explicit consistency rewards, like plan coherence and plan execution in MOCC-R1, or the reasoning becomes decoration.

5. Dropping audio from video pipelines. It's an easy cost cut, but it quietly re-creates the silent observer. Questions that hinge on sound will be guessed, and your benchmark probably won't catch it.

One thing to remember

All four papers converge on the same principle: the model shouldn't be the only judge of its own output. SSAV adds query agreement and spatial persistence. Video-HolmesV2 adds audio evidence and citation requirements. MASA adds a frozen external anchor. MOCC-R1 adds an auditable plan. In every case, the path to trust runs through constraints on what counts as evidence, not through a bigger model.

The Bottom Line

If you're deploying image captioning or VQA where a wrong object has real cost, such as medication assistance, accessibility, or navigation, add a training-free verification layer like SSAV before output. It needs no retraining, and the published results show roughly a third fewer sentence-level hallucinations.

If you're building long-video understanding, distrust accuracy numbers from visual-only benchmarks, and don't brute-force evidence with dense sampling. Use audio-anchored token compression and evaluate with evidence-based protocols, or you'll ship a silent observer that passes your test suite and fails in production.

If you're doing test-time adaptation under mixed distribution shifts, stop filtering by the model's own confidence. Anchor adaptation with a small set of frozen-MLLM semantic descriptions and keep the adaptive parameters few, or you're building a feedback loop that entrenches the original error.