acceptodds
Under review as a conference paper at ICLR 2027

LAM++: Enhanced LocateAnything via High-Resolution Post-Training and Geometry-Aware Speculative Decoding

Abstract

Recent vision-language models such as the LocateAnything Model (LAM) have enabled universal localization through generative parallel box decoding, yet they still have gaps with strong dedicated vision models such as SAM3 on dense scenes and small targets. We present **LAM++**, an enhanced LAM with higher accuracy and faster speed via post-training and speculative decoding. We first identify dense scenes and small targets as major bottlenecks, and construct a high-quality corpus of over 20M localization examples. A failure-driven SFT recipe that integrates high-resolution visual processing with single-query training is used to substantially improve the localization accuracy on all domains, and a sequence-level GRPO strategy is used to further strengthen performance. We then introduce a geometry-aware speculative decoding (GeoSpec) framework that reuses LAM's native parallel decoder (MTP) for drafting and employs the autoregressive (AR) policy for verification. Unlike exact token-level verification of speculative decoding, GeoSpec exploits the geometric structure of localization outputs, accepting spatially equivalent coordinate proposals under a relaxed criterion and invoking AR correction only when necessary. This yields a favorable accuracy–efficiency Pareto frontier: GeoSpec achieves a 68% speedup with ΔmF1 ≤ 0.1. Across a diverse suite of localization benchmarks that span different task formulations and visual domains, LAM++ achieves state-of-the-art performance, outperforming SAM3 as well as larger or closed-source VLMs. We will release our models, the full training corpus, and complete post-training recipes.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Object Grounding, VLM

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_ekWFOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes LAM++, an improved version of the LocateAnything Model (LAM) for language-guided object localization, focusing on dense scenes and small objects. The main contributions include: constructing a dataset of over 30 million localization examples, emphasizing challenging cases, using high-resolution images and single-query training to improve localization accuracy, applying GRPO with a localization-specific reward to improve bounding-box predictions, introducing geometry-aware speculative decoding (GeoSpec) that uses fast MTP predictions as drafts and AR predictions for verification. Experiments across 11 benchmarks show improvements over LAM and other localization models, particularly in dense scenes. GeoSpec achieves approximately 68% higher box-decoding throughput than AR decoding with small accuracy degradation.

Critical strengths
  • Failure-driven approach: The authors identify specific weaknesses of LAM in dense scenes and small-object detection, then design their training strategy to address these failures.
  • Strong empirical results: LAM++ substantially improves over the original LAM across different benchmarks, with particularly impressive improvements on Dense200.
  • Well-engineered paper: The paper combines large-scale data construction, high-resolution training, GRPO, and efficient decoding into a well-engineered localization framework.
  • Broad experimental evaluation: The experiments cover 11 benchmarks across object detection, dense localization, referring grounding, reasoning grounding, GUI grounding, and other domains.
  • Simple and effective GeoSpec: GeoSpec is well-motivated, straightforward to implement, and achieves significant decoding speed improvements with minimal accuracy loss.
Critical weaknesses
  • Limited methodological novelty: Most accuracy improvements rely on established techniques. High-resolution and single-query training are straightforward, failure-focused data selection is not conceptually new, and GRPO was previously used for generative localization in Rex-Omni. GeoSpec appears to be the main distinct algorithmic contribution.
  • Unclear contribution of additional computation: The model uses substantially higher image resolution, more training data, and extensive post-training. Comparisons with LAM are not fully matched in terms of resolution, training data, and computational budget.
  • Limited computational efficiency analysis: The paper emphasizes decoding throughput but lacks a sufficiently detailed comparison of overall inference latency, memory, and computational costs, particularly considering the higher input resolution.
  • Dependence on strong teacher models: The pseudo-labeling pipeline relies heavily on models such as SAM3 and Rex-Omni, raising questions about data quality, scalability, and how much of the improvement comes from stronger supervision.
  • Comparison with SAM3: The paper emphasizes outperforming SAM3, but the models have different architectures and intended capabilities. Some evaluated tasks may not be well suited to SAM3's query formulation, making parts of the comparison less informative.
  • Heuristic GeoSpec verification: GeoSpec uses a fixed coordinate tolerance, although the effect of coordinate errors depends on object size. An IoU-based or object-scale-aware acceptance criterion may be more appropriate, particularly for small objects.
Questions for authors
  • Q1. Compute-matched comparison: How much of the improvement over LAM comes from higher input resolution, additional training data, and training compute? Can the authors provide a more controlled comparison?

  • Q2. GeoSpec generalization: Can GeoSpec be applied to other autoregressive localization models that do not already have LAM's MTP policy? How much additional training or architectural modification would be required?

  • Q3. Data contamination: How do the authors ensure that the large-scale web-derived training data does not overlap with evaluation benchmarks?

  • Q4. GeoSpec acceptance: Have the authors considered IoU-based or object-size-aware acceptance criteria instead of a fixed coordinate tolerance? How does the current criterion affect small-object localization?

9 comments
AuthorsOpenReview ↗

Rebuttal Overview

We sincerely thank the reviewer for the careful reading and constructive suggestions. Such thoughtful feedback is particularly encouraging in today's fast-moving research environment and helps us further improve the paper and make a more meaningful contribution to the research community.

Before the point-by-point response, we briefly summarize the key revisions (see the General Response for details):

  • We re-evaluated all models using the same hardware and evaluation pipeline.
  • Every open-source model is now evaluated under four input settings (single/multi-query × original/high resolution), with Table 1, Fig. 4, and Tables A1–A7 updated accordingly.
  • The main conclusions remain unchanged. Notably, after full-scale training, GeoSpec outperforms AR decoding on 10 of 11 benchmarks, achieving an average gain of 1.6 mF1 while increasing decoding throughput by 68%, thus improving both accuracy and efficiency.
  • All six weaknesses are addressed point by point below, with the newly requested analyses consolidated in Appendix I (Tables A12–A15) of the revised PDF.
AuthorsOpenReview ↗

W1: Methodological novelty

We agree that high-resolution processing, single-query training, failure-focused selection, and GRPO are established, and we do not claim them as our inventions. Our contribution is the systematic diagnosis that motivates them, plus two results we believe are new:

  1. A quantified failure diagnosis that drives the recipe. LAM trails SAM3 by only 0.59 mF1 on long-tail LVIS queries but by 23.51 on dense queries (Table A8). The recipe is built against that measurement, and each component is then ablated separately (Part 2).
  2. Bidirectional cross-policy transfer. Optimizing only the AR policy improves MTP (62.48→63.85 LVIS mF1), and optimizing only MTP improves AR (64.16→65.71), even though neither receives the other's loss. MTP-GRPO, however, shifts both policies to a more conservative operating point (dense AR recall 38.24→36.09, vs. 42.01 for AR-GRPO). This characterizes RL transfer between decoding policies that share parameters.
  3. GeoSpec, which the reviewer already identifies as the main algorithmic contribution. We would add that its benefit is not only speed: accepting MTP's jointly predicted coordinates within a tolerance measurably improves box geometry, and the gain grows with the IoU threshold (LVIS: +1.0 F1 at IoU 0.5, +2.2 F1 at IoU 0.9).
AuthorsOpenReview ↗

W2+Q1: Compute-matched comparison with LAM

This is a fair concern, and we answer it at two levels.

(1) Inference Compute, Fully Matched (Table 1)

The input setting can be changed without any retraining, so we evaluated the original LAM under all four input settings. Mean mF1 over the 11 benchmarks:

Input SettingLAM ARLAM++ GeoSpecΔLAM++ Better On
Multi-query, original res. (LAM default)46.3350.16+3.848 / 11
Single-query, original res.47.8354.56+6.7411 / 11
Multi-query, high res.48.5653.21+4.659 / 11
Single-query, high res. (LAM++ default)49.2457.62+8.3811 / 11

Giving LAM the same high-resolution input closes only 2.24 mF1 of the 11.29 mF1 gap. Under every matched setting, including LAM's own default format which LAM++ was never trained for, LAM++ remains ahead. The advantage therefore does not come from spending more inference compute.

(2) Training Data and Compute (Table A8)

Each ablation step is matched on everything except the factor under study: RSL vs. SSL use the same 450K budget; single-query uses the same data as SSL; high resolution uses the same data and query format. This isolates:

  • +5.92 mF1 for the silver data;
  • +0.97 mF1 for single-query (+3.82 for MTP);
  • +10.02 mF1 for resolution;
  • +2.05 mF1 for AR-GRPO.
Reviewer_rX5kOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LAM++, an enhanced version of LAM for open-vocabulary localization. The authors construct a large-scale training corpus using existing models and improve LAM through high-resolution SFT and AR-GRPO. They also introduce GeoSpec, a geometry-aware speculative decoding strategy for faster inference. Experiments demonstrate improved localization accuracy across multiple benchmarks and a 68% increase in decoding throughput with negligible accuracy degradation.

Critical strengths
  1. Strong empirical performance. LAM++ achieves substantial improvements over LAM across multiple benchmarks, particularly on dense localization tasks.
  2. Efficient inference. GeoSpec improves decoding throughput by approximately 68% with minimal accuracy degradation, making autoregressive localization more practical.
Critical weaknesses
  1. Limited novelty and unclear sources of improvement. The method mainly combines existing techniques. The ablation results suggest that most accuracy gains come from high-resolution training, while GeoSpec primarily improves inference efficiency. The contributions of the expanded dataset and individual training strategies are not sufficiently isolated.
  2. Insufficient justification of key components. The quality of pseudo-labels generated by existing models is not quantitatively validated. Additionally, GeoSpec uses a fixed coordinate tolerance \epsilon, whose suitability for objects of different sizes remains unclear.
Questions for authors
  1. Performance attribution: Could the authors further disentangle the contributions of high-resolution training, the expanded dataset, and AR-GRPO? How much improvement comes from simply increasing the input resolution?
  2. Pseudo-label quality: How is the accuracy of pseudo-labels generated by SAM3 and RexOmni ensured? Could the authors report their precision and recall on a manually verified subset, particularly before and after merging?
  3. GeoSpec verification: Why is a fixed \epsilon=7 used for coordinate verification? Since the same coordinate error can have different effects on small and large objects, how sensitive is GeoSpec to this threshold across object scales?
7 comments
AuthorsOpenReview ↗

Rebuttal Overview

We sincerely thank the reviewer for the careful reading and constructive suggestions. Such thoughtful feedback is particularly encouraging in today's fast-moving research environment, and helps us further improve the paper and make a more meaningful contribution to the research community.

Before the point-by-point response, a brief note on the revision (details in the General Response):

  • Unified evaluation. We re-evaluated all models with the same hardware and the same evaluation pipeline.
  • Comprehensive input-setting evaluation. Every open-source model is now reported under four input settings (single/multi-query × original/high resolution), with Table 1, Fig. 4, and Tables A1–A7 updated accordingly.
  • Main conclusions preserved. The main conclusions are unchanged. One notable update: after full-scale training, GeoSpec is more accurate than AR decoding on 10 of 11 benchmarks (+1.6 mF1 on average). It thus improves accuracy while accelerating decoding.
  • Additional analyses. All new analyses requested in the reviews are collected in the new Appendix I (Tables A12–A15).

One update is directly relevant to this review. With full-scale training, GeoSpec is no longer only an efficiency method: it outperforms AR decoding on 10 of 11 benchmarks while also increasing decoding throughput by 68%. The advantage grows at stricter IoU thresholds (on LVIS, +1.0 F1 at IoU 0.5 vs. +2.2 F1 at IoU 0.9), indicating that geometry-aware verification refines box coordinates rather than merely approximating AR.

AuthorsOpenReview ↗

W1+Q1: Disentangling the Contributions, and How Much Comes from Resolution Alone (Part 1)

We agree that the individual techniques are established, and we report the contribution of each component explicitly rather than claiming them jointly.

(1) Controlled ablation on LVIS (Table A8). Each step changes exactly one factor, and we report both decoding policies of the same checkpoint:

StageLVIS AR mF1LVIS MTP mF1Δ AR
LAM48.2144.23–
+ silver labels, random (RSL, 450K)54.1350.25+5.92
+ silver labels, failure-driven (SSL, same 450K)53.1748.61−0.96
+ single-query training54.1452.43+0.97
+ high resolution64.1662.48+10.02
+ AR-GRPO (full LVIS)66.2163.85+2.05
+ AR-GRPO (LVIS-p, 40% of the data)65.8763.23−0.34

Key observations:

  • Expanded data contributes +5.92 mF1 on its own (LAM → RSL), independently of resolution.
  • Failure-driven selection is a redistribution rather than an increase: at the same 450K budget it improves the targeted subsets (dense 27.14→28.82, small 29.97→30.57) while slightly lowering the normal-case average, which dominates LVIS.
  • Single-query training helps AR modestly (+0.97) but MTP substantially (+3.82), narrowing the AR–MTP gap from 4.56 to 1.71. This matters because GeoSpec drafts with MTP.
  • High resolution is indeed the largest single factor (+10.02), and we state this plainly in Sec. 4.3; its gains concentrate exactly where we predicted (dense 29.80→43.16, small 31.19→47.13).
  • AR-GRPO adds +2.05 on top of the high-resolution SFT model, and also improves MTP by +1.37 without any direct MTP loss (cross-policy transfer).
AuthorsOpenReview ↗

W1+Q1: Disentangling the Contributions, and How Much Comes from Resolution Alone (Part 2)

(2) How much comes from simply increasing the resolution? (Table A14).

To answer this at full scale, we evaluated the original LAM under four input settings, so that the resolution and query format can be changed at inference time with no retraining. Mean mF1 over the 11 benchmarks:

Model / SettingMean mF1Δ vs. LAM default
LAM AR, default setting46.33–
LAM AR, best input setting per benchmark (no training)51.38+5.06
LAM++ AR, default setting56.06+9.74
LAM++ GeoSpec, default setting57.62+11.29

Simply giving the original LAM better inputs is worth +5.06 mF1, which is substantial and which we now report explicitly. It is, however, less than half of the +11.29 obtained by LAM++. Moreover, even when LAM is given its best input setting on every benchmark, LAM++ GeoSpec is still better on 11 of 11 benchmarks (+6.24 mF1 on average), and LAM++ remains better on 8 of 11 benchmarks even under LAM's own default setting, which LAM++ was never trained for.

(3) Each model family is reported under both of its decoding modes.

Table 1 reports the original LAM under AR and Hybrid decoding, and LAM++ under AR and GeoSpec decoding, on all 11 benchmarks. This makes the role of the fast decoding mode explicit, and the contrast is the clearest evidence that GeoSpec is not only an efficiency mechanism.

Mean mF1 over the 11 benchmarks, each model under its default setting:

FamilyAR DecodingFast Decoding ModeΔFast Mode Better On
LAM46.3343.52 (Hybrid)−2.812 of 11
LAM++56.0657.62 (GeoSpec)+1.5610 of 11

For the original LAM, decoding faster costs accuracy, which is exactly the accuracy–efficiency trade-off we set out to remove. For LAM++, decoding faster improves accuracy.

The same holds under every one of the four input settings (LAM Hybrid − AR ranges from −0.15 to −2.89; LAM++ GeoSpec − AR ranges from +0.30 to +1.56), so it is not an artifact of one protocol.

Reviewer_EUp4OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LAM++, an extension of the LocateAnything Model (LAM) for generative visual localization. The authors argue that LAM’s main failures are concentrated in dense scenes and small objects, and address them through four components: (i) constructing a large-scale localization corpus of over 30 million samples, covering detection, referring grounding, dense localization, reasoning grounding, and GUI grounding; (ii) failure-driven supervised fine-tuning combined with high-resolution visual processing and single-query training; (iii) AR-GRPO post-training, with a systematic study of GRPO on both autoregressive (AR) and multi-token prediction (MTP) policies; and (iv) GeoSpec, a geometry-aware speculative decoding scheme that uses LAM’s native MTP policy as the drafter and the AR policy as the verifier, replacing exact token agreement with coordinate tolerance. Experiments cover LVIS, SACo-Gold, COCO, Dense200, VisDrone, Talk2Car, HumanRef, RefCOCOg, RefBench-PRO, ReasonSeg, and ScreenSpot-Pro. The authors report that LAM++ outperforms SAM3 on 10 of 11 benchmarks and improves decoding throughput by approximately 68% over AR decoding with essentially unchanged accuracy.

Critical strengths
  • The paper does not merely claim that generative localization is insufficient; it identifies dense-scene instance enumeration and small-object granularity as the main bottlenecks through controlled analysis. This gives coherent motivation to the subsequent data, SFT, and RL design.
  • The work spans data construction, supervised training, reinforcement learning, and decoding efficiency. Compared with modifying a single module, this combination is more comprehensive and provides a useful recipe for generative localization.
  • Classical speculative decoding requires exact token agreement, but bounding-box coordinates are discrete tokens representing continuous geometry. Relaxing verification to a coordinate tolerance is well motivated, and the paper provides useful analyses of tolerance, target-size sensitivity, draft rejection, and suffix invalidation.
Critical weaknesses
  • The contribution is largely a systematic combination of existing techniques. High-resolution input, single-query training, GRPO, MTP drafting, and speculative decoding are all established ideas. The main methodological novelty is concentrated in GeoSpec, which is a natural extension of speculative decoding to geometric outputs. The paper should more clearly state what is conceptually new beyond integration and more carefully decompose the contribution of each component.
  • Data construction relies heavily on pseudo-labels and large/closed-source models, but quality and contamination analyses are insufficient. A large portion of the 30M corpus is generated or merged by Qwen3.5-35B-A3B, Qwen3.8-27B, SAM3, and RexOmni. The paper does not adequately report pseudo-label noise rates, cross-model agreement distributions, filtering sensitivity, or deduplication/contamination checks against evaluation benchmarks. For a paper claiming state-of-the-art localization, this is a critical omission. In particular, the risk that the silver-label pool overlaps with or is derived from evaluation distributions should be explicitly analyzed.
  • Comparisons with SAM3 and larger VLMs require stronger fairness controls. SAM3 is a dedicated localization/segmentation model, whereas LAM++ is a general VLM architecture. Comparison results depend on input resolution, inference budget, additional data, and tuning protocol. The claim that LAM++ outperforms SAM3 on 10 of 11 benchmarks may be strong, but the paper should clarify whether all models were evaluated under comparable input resolution, prompting, and post-processing, and should report the sensitivity of the conclusion to these factors.
  • GeoSpec remains sensitive to box ordering. The appendix acknowledges that GeoSpec directly verifies the serialized box sequence and does not perform set-level matching or reordering. If MTP and AR prefer different box orders, geometrically equivalent drafts may be rejected, reducing accepted length and speedup. The paper should quantify how often this occurs and whether a lightweight order-consistent drafting or order-invariant verification scheme can improve the accuracy–efficiency trade-off.
Questions for authors
  1. Please report pseudo-label noise estimates, cross-model agreement statistics, and the effect of filtering thresholds on downstream performance. Provide a contamination analysis between the silver-label pool and all evaluation benchmarks, especially LVIS, COCO, SACo-Gold, RefCOCOg, and ScreenSpot-Pro.
  2. Clarify the input resolution, prompting format, post-processing, and inference budget for SAM3 and all VLM baselines. If SAM3 was evaluated under a different protocol, state this explicitly and qualify the claim of outperforming SAM3 on 10 of 11 benchmarks accordingly.
  3. The negative results on structured action masking and fine-grained credit assignment are valuable. Please discuss more explicitly what they imply about the source of AR-GRPO’s gains and whether simpler RL formulations would suffice.
10 comments
AuthorsOpenReview ↗

Rebuttal Overview

We sincerely thank the reviewer for the careful reading and constructive suggestions. Such thoughtful feedback is particularly encouraging in today's fast-moving research environment, and helps us further improve the paper and make a more meaningful contribution to the research community.

Before the point-by-point response, a brief note on the revision (details in the General Response):

  • We re-evaluated all models with the same hardware and the same evaluation pipeline.
  • Every open-source model is now reported under four input settings (single/multi-query × original/high resolution), with Table 1, Fig. 4, and Tables A1–A7 updated accordingly.
  • The main conclusions are unchanged. One notable update: after full-scale training, GeoSpec is more accurate than AR decoding on 10 of 11 benchmarks (+1.6 mF1 on average). It thus improves accuracy while accelerating decoding.
  • All new analyses requested in the reviews are collected in the new Appendix I (Tables A12–A15).
AuthorsOpenReview ↗

W1: What is new beyond integration

We agree that high-resolution processing, GRPO, MTP, and speculative decoding exist individually, and we do not claim them as our inventions. Our contribution is twofold: a systematic diagnosis and controlled study of where VLM-based localization fails, and a decoding algorithm tailored to structured localization outputs. Concretely:

  1. Failure-driven post-training.
  • Rather than scaling generic detection data, we first diagnose where generative localizers fail: instance under-enumeration in dense scenes, and insufficient visual granularity for small objects. In Table A8, LAM trails SAM3 by only 0.59 mF1 on long-tail queries, but by 23.51 on dense queries.
  • We then build a 30.08M-example corpus and quantify, in controlled experiments, the separate effects of failure-driven data selection, single-query training, and high-resolution processing (Part 2).
  1. Bidirectional cross-policy transfer between AR and MTP. LAM's AR and MTP decoding policies share parameters but generate differently.
  • Optimizing only AR also improves MTP (LVIS MTP mF1 62.48→63.85).
  • Optimizing only MTP also improves AR (64.16→65.71).
  • However, MTP-GRPO moves the model to a more conservative precision–recall operating point: dense-subset AR recall drops from 38.24 (SFT) to 36.09, whereas AR-GRPO raises it to 42.01.

This characterizes how RL transfers across decoding policies that share parameters, which may be relevant to other models with MTP heads.

  1. GeoSpec: geometry-aware speculative verification. Standard speculative decoding requires exact token agreement, yet coordinate tokens encode continuous geometry, and different token sequences can denote nearly identical boxes. GeoSpec therefore:
  • verifies coordinate tokens with a geometric tolerance, while requiring exact agreement on structural tokens;
  • commits the longest valid prefix and lets AR complete the current box at the first rejection, so that later generation always conditions on a valid AR causal context;
  • reuses LAM's native MTP head as the drafter, with no extra draft model or training.

With full-scale training, GeoSpec outperforms AR on 10/11 benchmarks (+1.6 mF1 on average). Its advantage grows with the IoU threshold: on LVIS, it is +1.0 F1 at IoU 0.5 and +2.2 F1 at IoU 0.9. Accepting MTP's jointly predicted coordinates within the tolerance thus refines box geometry; GeoSpec is not merely an approximation of AR decoding.

AuthorsOpenReview ↗

W1: contribution of each component

(1) Controlled Ablation on LVIS (Fig. 6, Table A8)

Each step changes one factor:

StepChangeLVIS AR mF1Notes
LAM–48.21Starting point
SFT on RSL450K randomly sampled silver labels54.13+5.9 from silver labels
SFT on SSLSame 450K budget, failure-driven selection53.17vs. RSL: dense 27.14→28.82, small 29.97→30.57
+ Single querySame data, single-query format54.14MTP 48.61→52.43; AR–MTP gap 4.56→1.71
+ High resolutionSame data and query format64.16Largest gain: dense 29.80→43.16, small 31.19→47.13
+ AR-GRPO (LVIS-p)40% of the LVIS training data65.87On par with full-LVIS GRPO (66.21); better on dense/small

Takeaways:

  • Under the same data budget, failure-driven selection improves the targeted dense and small subsets, at a small cost on normal queries.
  • Single-query training improves both decoders and narrows the AR–MTP gap.
  • High resolution is the largest single factor, identifying visual granularity as LAM's main bottleneck.
  • For RL, a small set of hard samples is as effective as the full set.
  • AR-GRPO and MTP-GRPO reach similar mF1 (65.87 vs. 65.71), but AR-GRPO has clearly higher recall (65.18 vs. 62.15 overall; 42.01 vs. 36.09 on dense), at about 1/2.7 of the training cost.

(2) Full-Scale Attribution (Table 1; summarized in Table A14)

To separate the effect of the inference setting from that of training, we evaluated both LAM and LAM++ under all four settings. The table reports mean mF1 over the 11 benchmarks, with gains relative to LAM's default setting:

Model / SettingMean mF1vs. LAM default
LAM AR, default (multi-query, original resolution)46.33–
LAM AR, best setting per benchmark (no training)51.38+5.06
LAM++ AR (default setting)56.06+9.74
LAM++ GeoSpec (default setting)57.62+11.29

Takeaways:

  • Inference setting vs. training. Changing only the inference setting already helps LAM. But even when LAM uses its best setting on every benchmark, LAM++ GeoSpec is better on 11/11 benchmarks (+6.2 mF1 on average), so training contributes gains that the inference setting cannot replace.
  • LAM's own default. Under LAM's default setting, a format LAM++ was not trained on, LAM++ still outperforms LAM on 8/11 benchmarks.
  • Input format should match training format.
    • Models trained with multi-query inputs (LAM, RexOmni) degrade with single-query inputs on most detection benchmarks.
    • LAM++, trained with single queries, degrades with multi-query inputs.
    • For the general-purpose Qwen3.8-27B, single-query inputs help on every benchmark that contains multi-query cases. With single-query inputs, high resolution helps or is neutral on all benchmarks.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.