MatchTIR: Advanced Matching Frameworks

• MatchTIR is a family of frameworks that applies mathematically principled matching algorithms—such as bipartite matching, optimal transport, and stable marriage—to enhance fine-grained supervision.
• It refines credit assignment in tool-integrated reasoning by matching predicted and ground-truth tool calls, yielding denser, more accurate reward signals and improved overall performance.
• In IRS-aided THz networks and TIR image analysis, MatchTIR achieves near-optimal associations and state-of-the-art performance while maintaining scalable computational complexity.

MatchTIR refers to a family of frameworks that apply matching-based algorithms and theory—often bipartite or multi-object matching—to enhance fine-grained supervision and association in domains as varied as tool-integrated reasoning with LLMs, terahertz communications with reconfigurable surfaces, and thermal-infrared (TIR) image matching and verification. The core concept is the application of mathematically principled matching or assignment (such as bipartite matching, optimal transport, or stable marriage algorithms) to produce more informative, granular credit signals or associations than previous global or undifferentiated approaches.

1. Motivation and High-Level Problem Formulation

In tool-integrated reasoning (TIR) with LLMs, the agent must interleave multi-turn natural language reasoning with calls to external tools (APIs, calculators, search engines, etc.), constructing stepwise trajectories

$T = (s_1, s_2, \ldots, s_T)$

where each $s_t$ includes (i) a reasoning step, (ii) one or more tool calls, and (iii) the resulting observations. Standard reinforcement learning (RL) approaches assign outcome- or trajectory-level rewards uniformly across all turns. This coarse credit assignment cannot distinguish effective, redundant, or erroneous tool calls, leading to suboptimal policy learning, particularly in long-horizon, multi-step tasks where early errors have cascading impacts (Qu et al., 15 Jan 2026).

In other domains, such as IRS-aided THz communications, the challenge is to associate sources, reflective elements, and destinations in a large combinatorial space. The global S–I–D (Source–IRS–Destination) association problem is NP-hard if tackled via exhaustive search (Rahim et al., 2024).

Across these use cases, MatchTIR-style frameworks introduce fine-grained matching as a solution to the limitations of existing global or heuristic approaches.

2. Matching-Based Credit Assignment and Association

Tool-Integrated Reasoning

MatchTIR (Qu et al., 15 Jan 2026) introduces a bipartite matching-based formulation for assigning rewards at the level of individual tool calls:

• Predicted calls $P = \{p_1, \dots, p_m\}$ are matched to ground-truth calls $G = \{g_1, \dots, g_n\}$ from annotated trajectories.
• A similarity score $s(i,j) \in [0,1]$ is computed for each pair, incorporating:
  1. Tool name coincidence,
  2. Jaccard similarity of parameter names,
  3. Normalized content match between parameter values.

• Two assignment strategies:

  • Hard assignment: Solve the maximum-weight bipartite matching (Hungarian/Kuhn-Munkres algorithm), rewarding only exact matches and penalizing spurious/unmatched predictions.
  • Soft assignment: Relaxed one-to-many assignment via discrete optimal transport (OT), yielding fractional credit for near-matches.

For turns issuing multiple tool calls, the reward for the turn is the mean of its call-level rewards. This yields turn-level dense supervision.

IRS-Aided Terahertz Networks

In (Rahim et al., 2024), the goal is optimal user association in the context of IRS-aided THz wireless networks. MatchTIR here refers to the matching-theoretical solution for the S–I–D assignment problem:

• The sum-rate maximization objective leads to a 3D-matching NP-hard problem.
• The approach decomposes the problem into two consecutive stable-matching subproblems, each solved with a Gale–Shapley (deferred-acceptance) procedure:
  1. Source–IRS matching (P1),
  2. (Source–IRS)–Destination matching (P2).

• This framework guarantees solution stability (no blocking pairs) and achieves near-optimal rates at polynomial computational cost. Explicit channel models and cascaded pathloss formulae are used to compute utility matrices used in the matching (Rahim et al., 2024).

TIR Matching for Image Analysis and Tracking

In visible–TIR image matching, e.g., XoFTR (Tuzcuoğlu et al., 2024), coarse-level assignment leverages adaptive, dual-softmax similarity matrices to generate initial candidate correspondences, followed by multi-scale refinement and sub-pixel regression. The final assignment can be seen as an adaptive matching pipeline, with one-to-one or one-to-many assignments reflecting local feature reliability and scale variance.

In verification-style TIR tracking (HSSNet (Li et al., 2017)), the problem is cast as determining which candidate window most closely matches a target template, via a Siamese network whose response map reflects a matching score at each location.

3. Dual-Level Advantage and Reward Aggregation

In tool-integrated reasoning, MatchTIR combines local (turn-level) and global (trajectory-level) reward signals into an integrated advantage term used by the policy optimizer (Qu et al., 15 Jan 2026):

• Turn-level advantages: Estimate the local contribution of a turn across rollouts, normalizing discounted future rewards at step $t$.
• Trajectory-level advantages: Capture overall task success, aggregated over the full rollout.
• Integrated dual-level advantage:

$A^{dual}_{i,j} = A^{traj}_i + A^{turn}_{i,t}$

or more generally as a convex combination. Distinct values per token and turn ensure that only impactful actions (tool calls) receive strong gradients during learning.

4. Empirical Results and Practical Performance

Tool-Integrated Reasoning

Experiments on three benchmarks (FTRL, BFCL, ToolHop) with model backbones Qwen3-4B/8B (Qu et al., 15 Jan 2026) demonstrate:

• MatchTIR outperforms baseline RL approaches by +4.7–4.8 points in Solve-F1 or final answer accuracy.
• The 4B model with MatchTIR surpasses many 8B competitors, especially on long-horizon tasks demanding many tool calls.
• With increasing trajectory complexity (#tool calls per query), gains increase (e.g., +6.4 points on "hard" Qwen3-8B subsets).
• Learned agents invoke fewer tools on average, with higher correctness (increase of 8.2 points in successful call rate).

IRS-Aided THz Networks

In simulation, MatchTIR (matching-based association) achieves identical or near-identical sum-rates versus partial or full exhaustive search (within <1% gap), with up to 30% improvement over random or heuristic assignment at high transmit power. Complexity scales as $O(KN+LN)$ for $K$ sources, $N$ IRSs, $s_t$0 destinations (Rahim et al., 2024).

TIR Image Matching and Tracking

In TIR–RGB matching, the XoFTR pipeline leveraging adaptive, scale-aware assignment and sub-pixel refinement yields state-of-the-art pose estimation AUCs (e.g., 22.03% @5°, 39.03% @10°, 55.06% @20° on the METU-VisTIR benchmark), surpassing LoFTR and DKM by large margins (Tuzcuoğlu et al., 2024).

HSSNet for tracking demonstrates that direct matching-style training and localization, coupled with spatial-aware modules and channel attention, drives robust performance under occlusion, deformation, and large appearance changes (Li et al., 2017).

5. Limitations, Variants, and Extensions

Limitations

• MatchTIR in RL-for-TIR depends on the availability of verifiable ground-truth tool traces. In open-ended or creative settings, a unique reference trajectory may not exist, complicating reward construction (Qu et al., 15 Jan 2026).
• Computational cost increases due to bipartite matching (Hungarian/OT) and group-based advantage estimation.
• In IRS-based networks, the framework assumes channel reciprocity and complete utility matrices for matching; practical deployment scenarios may introduce partial observability.
• For TIR–RGB matching, current augmentation approaches are heuristic, and sub-pixel localization incurs modest runtime overhead (Tuzcuoğlu et al., 2024).

Variants and Future Directions

• RL: Transformer-based value estimation to reduce variance; "soft" reward models via signal distillation to handle multiple plausible traces; scaling matching frameworks for hundreds of APIs and more complex tool schemas (Qu et al., 15 Jan 2026).
• THz: More complex association models accounting for multi-user interference, mobility, or partial information (Rahim et al., 2024).
• TIR matching: End-to-end joint learning of TIR synthesis and matching (GAN-based), temporal/sequential modeling for video, and extensions to other sensor-modal pairs (Tuzcuoğlu et al., 2024).

6. Summary Table of MatchTIR Implementations

Domain	Matching Type	Objective	Key Method/Metric
Tool Reasoning	Bipartite (exact/OT)	Turn-level reward assignment	Hungarian/OT; Solve-F1
IRS THz Networks	Stable matching	S–I–D association (sum-rate)	Gale–Shapley; sum rate
TIR–RGB Image	Adaptive (dual-softmax, coarse-to-fine)	Feature correspondence	Pose AUC, corner error
TIR Tracking	Location matching	Target localization	Siamese response; EAO, accuracy

This table encapsulates the architectural and methodological diversity of MatchTIR across application domains. Bipartite and multi-object matching serve as a unifying principle for enhancing fine-grained credit assignment, association, or verification, yielding empirical gains in each respective area.

7. Concluding Insights

The MatchTIR paradigm demonstrates that incorporating mathematically principled matching into the association or credit assignment step—rather than relying on global or ad-hoc signals—enables substantially finer policy shaping, more precise feedback, and higher downstream success rates. In tool-integrated reasoning, this yields RL-trained LLMs that call fewer tools with higher correctness, rivaling much larger models. In wireless and computer vision, matching-based association and verification outperform heuristics and classifiers, efficiently exploiting the combinatorial structure of the problem. The principal limitations are reliance on good reference traces/utility signals and computational overhead at scale. Future trends point toward integrating learned, modality-agnostic reward or similarity models and expanding the scalability and generality of the matching-theoretical architecture.