The one-paragraph version
INTCORT is a training-free trick for making Vision-Language Models, meaning systems that take an image plus a text question and generate a text answer, better at simple spatial questions like “is the keyboard left of the notebook?” without retraining them or rewiring their insides. The authors start from a simple observation: when a model gets left/right or above/below wrong from the original view, it will often get it right if you flip the image or swap the names in the question, and when it is right it tends to be more confident about the word that names the relation. INTCORT therefore asks the same spatial question several times — once on the original image and question, once on a horizontally flipped image, once on a vertically flipped image, and once with the question reversed — maps all answers back into the original left/right language, and then lets the branches vote, with more confident branches getting more weight. Across seven spatial benchmarks and six open models it consistently beats the base model and prior training-free methods, with a reported average gain of 10.01% over base models, while leaving non-spatial performance essentially unchanged.
The problem they're solving, and why it matters
The problem is spatial reasoning, meaning the ability to describe where things are in an image in words, for example whether object A is to the left of, to the right of, above, or below object B, or where a single object sits relative to the picture itself. Despite strong progress on captioning and general question answering, current Vision-Language Models remain unreliable on even these basic relations. That matters because robots, navigation systems, and scene-understanding tools cannot safely act on “pick up the mug to the left of the knife” if left and right are flipped half the time.
Prior fixes fall into two camps, both with costs. Training-dependent methods, meaning methods that fine-tune the model on large new spatial datasets, are expensive and risk catastrophic forgetting, meaning the model learns the new spatial skill but forgets or degrades older abilities. Training-free methods that edit internals, meaning methods that intervene in attention patterns or decoding such as AdaptVis, DoLa, and VCD, avoid retraining but tamper with the mechanism the model uses for everything else, which can hurt general capabilities. The paper asks: can we improve spatial reliability purely from the outside, at input time, without touching weights, attention, or decoding?
The key idea, in plain words
The key idea is to treat one spatial question as multiple independent witnesses rather than one guess.
Imagine you are unsure whether the keyboard is left or right of the notebook in a photo. You might flip the photo horizontally in your head, you might turn it upside down, and you might rephrase the question as “where is the notebook relative to the keyboard?” If most of those reframed views point to the same underlying arrangement, and especially if the confident ones agree, you trust that answer more than your first uncertain guess.
INTCORT does that literally. A geometric image transformation, meaning a simple flip of the pixels such as mirror left-to-right or top-to-bottom, gives the model a new visual view of the same physical layout. A query-reversal transformation, meaning swapping the target and reference objects in the wording — for example changing “Where is the mug in relation to the knife?” to “Where is the knife in relation to the mug?” — gives the model a new linguistic view where the correct answer must flip in a predictable way. A relation-token confidence, meaning how strongly the model predicted the answer word itself such as “left,” is used to decide which views to trust. Correct views, the authors hypothesize, will on average speak more loudly than mistaken ones.
How it actually works, step by step
Step 0: Decide if this is even spatial. Not every question should be flipped. Flipping the image for “Is there a banana?” would just add noise. So INTCORT first sends the text question alone to a small language classifier, in the experiments Qwen2.5-1.5B-Instruct, which outputs one of two labels: spatial or non-spatial. Non-spatial inputs skip INTCORT entirely and go straight to the main Vision-Language Model unchanged. This is the preservation mechanism.
Step 1: Build multiple inference branches. For a question judged spatial, INTCORT creates a small set of inputs. One branch is the direct branch, meaning the original image and original question with no change. The others use a transformation set made of two families. The image family in the paper is fixed to two flips: H for horizontal flip and V for vertical flip. The query family is reversal with one variant, so N equals 1 in this implementation. That gives four branches total in practice: original, horizontally flipped, vertically flipped, and query-reversed.
A concrete example helps. Suppose the original truth is “keyboard is right of notebook.” On a horizontally flipped image, right becomes left visually, so if the model answers “left” on that flipped image, after mapping back that counts as evidence for “right” in the original space. Similarly, if the reversed question “Where is the notebook relative to the keyboard?” yields “left,” that also maps back to “right” for the original question.
Step 2: Get an answer and a confidence from each branch. Each branch is run independently through the same frozen Vision-Language Model, which generates a response. From each response INTCORT extracts the predicted relation label, meaning the normalized spatial word such as left, right, above, or below, and then maps it back to the original answer space as described above.
It also computes a confidence score for that label. In words: look only at the tokens that spell out the relation word, take the model’s log-probability for each of those tokens given the image, question, and prior words, and keep the weakest link — the minimum. The intuition is that a two-word answer is only as trustworthy as its least certain word. In symbols, after that verbal definition, the paper writes this confidence for branch k as the minimum over its relation-token positions of the log prediction probability, then turns it into a positive vote weight by exponentiating it.
Step 3: Weighted vote. For each possible relation, add up the weights of all branches whose mapped-back prediction equals that relation. Branches that agree pool their evidence, and higher-confidence branches contribute larger weights. The final prediction is simply the relation with the largest total. In words before symbols: total support for “left” equals sum of weights of branches voting left after mapping; pick the max.
What they showed — results, honestly, with limitations
The authors test on seven spatial benchmarks covering clean toy scenes with two objects on plain backgrounds and real photos from MS COCO, Visual Genome, and CV-Bench: Controlled-A with one large and one small object, Controlled-B with two small objects, COCO-one for single-object position versus the image, COCO-two for two-object relations, VG-one and VG-two similarly built from Visual Genome, and CV-Bench-R for diverse pairwise relations. They test six models: Qwen-VL-Chat, Qwen2-VL-7B, Qwen2.5-VL-7B, Qwen3-VL-8B, Molmo-7B, and LLaVA-NeXT-7B, against AdaptVis, DoLa, and VCD, using accuracy.
First, they validate the two hypotheses on Qwen2-VL-7B on Controlled-B plus CV-Bench-R. Of 141 samples wrong by direct inference, 130 were recovered by at least one transformed view, a recovery ratio of 92.20%, and in 107 of those 130 the correct transformed prediction had higher relation-token confidence than the wrong direct prediction, a confidence-superiority ratio of 82.31%. This is striking but narrow: it is shown only for one model on two datasets, so it is unclear how well it generalizes.
Second, INTCORT is best or tied-best on all benchmarks for five of the six models, with an average gain reported as 10.01% over base models and 6.47% over AdaptVis, averaged over those five models and all seven benchmarks. Gains can be large where base spatial ability is poor — for example Qwen-VL-Chat on VG-two rises from 36.05% to 57.94% — and smaller where the base model is already near ceiling, such as Qwen3-VL-8B on Controlled-A staying at 99.40%. The abstract also claims up to 25.01% improvement over prior works, but the paper body does not clearly anchor that maximum to a specific model-benchmark cell, so treat it as an author claim rather than a verified breakdown.
On the sixth model, LLaVA-NeXT-7B, INTCORT beats the base by 10.52% on average and beats AdaptVis on most benchmarks, but loses to AdaptVis on a few, notably Controlled-A where AdaptVis jumps to 97.29% versus 53.11% for INTCORT alone. The authors suggest INTCORT relies on latent ability that LLaVA-NeXT lacks, while AdaptVis more forcefully rewires attention. Importantly, INTCORT plus AdaptVis together is best on LLaVA-NeXT across all seven sets, supporting the claim that input-side voting and internal attention editing are complementary. On non-spatial checks with Qwen2-VL-7B, POPE stays identical and MME moves only from 88.08% to 88.12%, suggesting no general degradation, though this is shown for only one model and two benchmarks. Ablations show the larger 1.5B screening classifier beats a 0.5B version and that minimum-token confidence beats mean or max pooling on the two tested settings.
Why this might matter to me
If you care about spatial reasoning in LLMs and VLMs, INTCORT offers a cheap inference-time wrapper you can put around any frozen model to squeeze out more reliable left/right/above/below judgments for grounding, referring expressions, and embodied prompts, without risking the forgetting or interference that comes with fine-tuning or attention surgery — and you can still stack it with those methods later.
For world models and 3D generation and scene representation, the lesson is that 2D view consistency is a useful proxy for 3D coherence: asking the same relational question from mirrored views and requiring agreement is conceptually similar to enforcing multi-view consistency in NeRFs, Gaussian splats, or video world models. You could imagine extending the transformation set beyond flips to rendered novel views, depth-aware crops, or camera rotations, and using the same confidence-weighted vote as a self-check for layout, occlusion, or contact.
For text-to-CAD and parametric geometry, where a phrase like “boss to the left of the hole” must compile to exact constraints, a training-free reliability layer is attractive because you cannot afford to degrade the code-generation ability of the base model. INTCORT suggests a pattern: generate candidate constraints from multiple canonicalized sketches or viewpoints, map them back to a shared parameter space, and let token confidence arbitrate — a lightweight way to reduce flipped mates and misplaced features before invoking an expensive solver or verifier.
Terms worth knowing
Vision-Language Model: a system that conditions text generation on both an image and a text prompt to answer questions or describe scenes.
Spatial reasoning: here, the narrow task of naming where entities are, either single-entity position versus the image or pairwise relations like left, right, above, below between two objects.
Training-free: improving performance without updating model weights, by changing inputs, prompts, or how outputs are selected.
Catastrophic forgetting: loss of previously learned abilities after fine-tuning on new data.
Geometric image transformation: a pixel-level change that preserves the scene but alters viewpoint language, here horizontal flip and vertical flip.
Query-reversal transformation: swapping target and reference in the question, e.g. mug-relative-to-knife becomes knife-relative-to-mug, so the expected answer inverts predictably.
Relation-token confidence: how likely the model thought the spatial answer word itself was, measured from the generation probabilities of just those tokens.
Confidence routing / weighted aggregation: letting each transformed branch vote for its mapped-back answer with strength proportional to its confidence, then picking the answer with the most total weight.
Direct branch: the baseline inference run on the original untransformed image and question, included as one voter alongside transformed branches.