TQP in Modern Research Domains

• TQP is a multifaceted term that defines key technical constructs across compression, LLM prompting, action recognition, video streaming, databases, and optimization.
• It underpins specific methodologies such as texture quantization parameters for point-cloud compression, transition quantifying prompts for LLMs, and tensor query processors in database systems.
• TQP also drives advancements in temporal query projection, transcoding quality prediction, and quadratic optimization in massive-MIMO beamforming and integer semidefinite programming.

TQP is a context-dependent acronym that denotes several distinct technical constructs in recent research. In the cited literature, it refers to a Texture Quantization Parameter in point-cloud compression, a Transition Quantifying Prompt in prompt-based analysis of LLMs, a Temporal Query Projection module for skeleton-based action recognition, Transcoding Quality Prediction in adaptive video streaming, a Tensor Query Processor for compiling SQL into tensor programs, a Trace Quotient Problem in massive-MIMO beamforming, a Ternary Quadratic Problem in integer semidefinite optimization, and TQ-PreTraining for training a 120B sparse mixture-of-experts model (Long et al., 2024, Sato, 16 Apr 2025, Ye et al., 13 Nov 2025, Menon et al., 2023, Asada et al., 2022, He et al., 2022, Kim et al., 2014, Meijer et al., 30 Mar 2026, Shravan, 5 Jun 2026).

1. Terminological scope

The acronym is best understood through its local expansion and disciplinary setting. The same three letters denote objects of very different types: scalar codec parameters, prompt templates, projection modules, system architectures, optimization problems, and training regimes.

Expansion of TQP	Domain	Representative source
Texture Quantization Parameter	Point-cloud compression and PCQA	(Long et al., 2024)
Transition Quantifying Prompt	LLM behavioral analysis	(Sato, 16 Apr 2025)
Temporal Query Projection	Skeleton-based action recognition	(Ye et al., 13 Nov 2025)
Transcoding Quality Prediction	Adaptive video streaming	(Menon et al., 2023)
Tensor Query Processor	Database systems and tensor runtimes	(Asada et al., 2022, He et al., 2022)
Trace Quotient Problem	Massive-MIMO beamforming	(Kim et al., 2014)
Ternary Quadratic Problem	Integer semidefinite optimization	(Meijer et al., 30 Mar 2026)
TQ-PreTraining	Sparse MoE training	(Shravan, 5 Jun 2026)

This range of meanings is itself a useful classificatory fact. In the cited works, TQP may denote a control variable, an induced measurement function, a neural projection block, a query execution engine, a ratio-form optimization problem, or an exact problem class over $\{0,\pm 1\}^n$.

2. Media coding, bitrate control, and perceptual quality

In point-cloud compression, TQP denotes Texture Quantization Parameter. Under MPEG G-PCC in the Trisoup-Lifting mode, it is the quantization parameter used for texture attributes, analogous to video QP: low TQP yields fine quantization and higher texture fidelity, while high TQP yields coarse quantization, more texture distortion, and lower attribute bitrate (Long et al., 2024). The paper fixes geometry as lossless when calibrating the texture-only model and encodes each reference point cloud with

$\mathrm{TQP} \in \{28,34,40,46,51\}.$

Its no-reference bitstream-layer model, streamPCQ-TL, links TQP to texture bitrate per point (TBPP), texture complexity (TC), and subjective quality. For fixed TQP, TC is modeled as linear in TBPP,

$$\mathrm{TC}=s(\mathrm{TQP})\cdot \mathrm{TBPP}+i(\mathrm{TQP}),$$

with

$$s(\mathrm{TQP})=a_1\mathrm{TQP}^2+a_2\mathrm{TQP}+a_3,\qquad i(\mathrm{TQP})=b_1\mathrm{TQP}+b_2,$$

and the texture-only MOS is modeled as

$$\mathrm{MOS}_T=(\alpha\cdot \mathrm{TC}+\beta)\cdot \mathrm{TQP}+b.$$

The final no-reference estimate multiplies this texture term by a geometry attenuation factor $D_G(\mathrm{tNSL})$, a function of trisoupNodeSizeLog2, reflecting the empirical observation that heavy geometry degradation can dominate perceived quality (Long et al., 2024).

In adaptive video streaming, TQP denotes Transcoding Quality Prediction, formalized by the Transcoding Quality Prediction Model (TQPM). The task is to predict the post-transcoding visual quality of a segment without access to the final reconstructed video, using reduced-reference features extracted from the input segment together with the target bitrates of each transcoding stage (Menon et al., 2023). The model uses three DCT-energy-based features on the luma channel—average texture energy, average temporal energy, and average luminescence—computed over $T=8$ chunks per 4-second segment, and augments each chunk feature vector with the bitrate vector $B=[b_1,\dots,b_M]$. A single-layer LSTM with 50 cells predicts PSNR, SSIM, and VMAF. For single-stage transcoding, the reported average $R^2$ values are $0.83$, $\mathrm{TQP} \in \{28,34,40,46,51\}.$0, and $\mathrm{TQP} \in \{28,34,40,46,51\}.$1, with MAE $\mathrm{TQP} \in \{28,34,40,46,51\}.$2 dB, $\mathrm{TQP} \in \{28,34,40,46,51\}.$3 dB, and $\mathrm{TQP} \in \{28,34,40,46,51\}.$4, respectively; for two-stage transcoding, the averages are $\mathrm{TQP} \in \{28,34,40,46,51\}.$5, $\mathrm{TQP} \in \{28,34,40,46,51\}.$6, and $\mathrm{TQP} \in \{28,34,40,46,51\}.$7, with MAE $\mathrm{TQP} \in \{28,34,40,46,51\}.$8 dB, $\mathrm{TQP} \in \{28,34,40,46,51\}.$9 dB, and $$\mathrm{TC}=s(\mathrm{TQP})\cdot \mathrm{TBPP}+i(\mathrm{TQP}),$$0 (Menon et al., 2023). The average processing time is $$\mathrm{TC}=s(\mathrm{TQP})\cdot \mathrm{TBPP}+i(\mathrm{TQP}),$$1 s for a 4-second segment.

Taken together, these two usages place TQP inside rate–distortion modeling, but at different layers. In the point-cloud paper it is a direct codec-side control variable; in the streaming paper it is the name of the prediction task itself. This suggests that in media systems the acronym often appears near bitrate-control, reduced-reference, or perceptual-quality modeling, but its local expansion remains essential.

3. Prompting, sequence projection, and model scaling

In prompt-based LLM analysis, TQP denotes Transition Quantifying Prompt. The framework in "Waking Up an AI" is explicitly two-part: a Transition-Inducing Prompt (TIP) perturbs the target model, and a Transition Quantifying Prompt (TQP) is then given to a separate evaluator LLM, which scores the target response on structured affective dimensions (Sato, 16 Apr 2025). The evaluator is asked to provide a Tone Phase Classification, a Tsun-Dere Score (0–10), Emotive Markers Detected, a Phase Shift Point (if applicable), Quote-Based Evidence, and an Overall Interpretation. Quantitatively, the paper treats Tone Phase and Tsun-Dere as the main endpoints and compares prompt conditions with one-tailed Welch’s $$\mathrm{TC}=s(\mathrm{TQP})\cdot \mathrm{TBPP}+i(\mathrm{TQP}),$$2-tests. For GPT-4o as evaluator, the design yields $$\mathrm{TC}=s(\mathrm{TQP})\cdot \mathrm{TBPP}+i(\mathrm{TQP}),$$3 evaluations per TIP condition; for Gemini 2.5Pro, $$\mathrm{TC}=s(\mathrm{TQP})\cdot \mathrm{TBPP}+i(\mathrm{TQP}),$$4 per TIP condition. The main result is negative with respect to semantic fusion: TIPe and TIPn-e receive significantly higher Tone Phase and Tsun-Dere scores than the logical control TIPc, but TIPe vs TIPn-e is not significant (Sato, 16 Apr 2025).

In skeleton-based action recognition, TQP denotes Temporal Query Projection. In SUGAR, long skeleton sequences are first encoded by a CTR-GCN-based skeleton encoder and then passed through TQP before entering LLaMA2-7B (Ye et al., 13 Nov 2025). TQP is built from Q-Former blocks and a linear layer. Given a skeleton representation $$\mathrm{TC}=s(\mathrm{TQP})\cdot \mathrm{TBPP}+i(\mathrm{TQP}),$$5, the method defines a learnable query vector $$\mathrm{TC}=s(\mathrm{TQP})\cdot \mathrm{TBPP}+i(\mathrm{TQP}),$$6 and processes temporal segments sequentially, reusing the previously queried representation as the next query: $$\mathrm{TC}=s(\mathrm{TQP})\cdot \mathrm{TBPP}+i(\mathrm{TQP}),$$7 The parameters of the Q-Former are shared across segments. Empirically, on Toyota Smarthome, the reported accuracies are $$\mathrm{TC}=s(\mathrm{TQP})\cdot \mathrm{TBPP}+i(\mathrm{TQP}),$$8 for X-Attn., $$\mathrm{TC}=s(\mathrm{TQP})\cdot \mathrm{TBPP}+i(\mathrm{TQP}),$$9 for one Q-Former, $$s(\mathrm{TQP})=a_1\mathrm{TQP}^2+a_2\mathrm{TQP}+a_3,\qquad i(\mathrm{TQP})=b_1\mathrm{TQP}+b_2,$$0 for one linear layer, and $$s(\mathrm{TQP})=a_1\mathrm{TQP}^2+a_2\mathrm{TQP}+a_3,\qquad i(\mathrm{TQP})=b_1\mathrm{TQP}+b_2,$$1 for Temporal Query Projection (Ye et al., 13 Nov 2025). A token-length study further leads the paper to define the query length $$s(\mathrm{TQP})=a_1\mathrm{TQP}^2+a_2\mathrm{TQP}+a_3,\qquad i(\mathrm{TQP})=b_1\mathrm{TQP}+b_2,$$2 as $$s(\mathrm{TQP})=a_1\mathrm{TQP}^2+a_2\mathrm{TQP}+a_3,\qquad i(\mathrm{TQP})=b_1\mathrm{TQP}+b_2,$$3, since reducing sequence length from $$s(\mathrm{TQP})=a_1\mathrm{TQP}^2+a_2\mathrm{TQP}+a_3,\qquad i(\mathrm{TQP})=b_1\mathrm{TQP}+b_2,$$4 to $$s(\mathrm{TQP})=a_1\mathrm{TQP}^2+a_2\mathrm{TQP}+a_3,\qquad i(\mathrm{TQP})=b_1\mathrm{TQP}+b_2,$$5 improves performance, whereas compressing to length $$s(\mathrm{TQP})=a_1\mathrm{TQP}^2+a_2\mathrm{TQP}+a_3,\qquad i(\mathrm{TQP})=b_1\mathrm{TQP}+b_2,$$6 substantially degrades it.

In large-scale sparse-MoE training, TQP denotes TQ-PreTraining. In "Reversible Foundations," the 120B LightningLM stage is trained through a strategy that combines quantized base expert weights with trained low-rank adapters, while carrying optimizer state only on the adapter parameters (Shravan, 5 Jun 2026). The paper states that the 120B model has 20 layers, 460 routed experts per layer, and top-12 routing, with more than 100B parameters in the routed experts alone. TQP addresses the optimizer-memory wall by holding the expert base weights in an 8-bit TurboQuant representation and adding rank-16 adapters. The expert-path optimizer state is thus carried on 2.26B adapter parameters rather than the more than 100B parameters resident in the routed experts, reducing expert-path optimizer state by roughly a factor of $$s(\mathrm{TQP})=a_1\mathrm{TQP}^2+a_2\mathrm{TQP}+a_3,\qquad i(\mathrm{TQP})=b_1\mathrm{TQP}+b_2,$$7 (Shravan, 5 Jun 2026). The final 120B lineage is reported as flushless TQP; the released checkpoint reaches a trailing-100 training loss of $$s(\mathrm{TQP})=a_1\mathrm{TQP}^2+a_2\mathrm{TQP}+a_3,\qquad i(\mathrm{TQP})=b_1\mathrm{TQP}+b_2,$$8.

Across these three machine-learning usages, TQP serves three different roles: a measurement prompt, a temporal compression bridge, and a memory-efficient training regime. The commonality is procedural rather than semantic: each TQP mediates between a high-dimensional process and a compact, analyzable interface.

4. Tensor-native query processing

In database systems, TQP denotes Tensor Query Processor, pronounced “teacup” in the 2022 demonstration paper (Asada et al., 2022). The system compiles relational queries into tensor programs and runs them on tensor computation runtimes such as PyTorch, ONNX Runtime, TVM, and PyTorch/XLA (Asada et al., 2022, He et al., 2022). Its architecture has four layers—Parsing, Optimization, Planning, and Execution—with Spark used as the current front-end for SQL parsing and physical-plan generation. The internal representation is then lowered to tensor programs implementing relational operators by combinations of tensor primitives such as sort, bincount, cumsum, bucketize, scatter, masked_select, unique, and unique_consecutive (He et al., 2022).

TQP uses a columnar tensor representation in which each logical column is a 2-D tensor of shape $$s(\mathrm{TQP})=a_1\mathrm{TQP}^2+a_2\mathrm{TQP}+a_3,\qquad i(\mathrm{TQP})=b_1\mathrm{TQP}+b_2,$$9. Numeric and date columns are encoded as $$\mathrm{MOS}_T=(\alpha\cdot \mathrm{TC}+\beta)\cdot \mathrm{TQP}+b.$$0 tensors, while strings are represented as fixed-width character matrices, allowing predicates such as LIKE and equality to be expressed as tensor operations (Asada et al., 2022, He et al., 2022). The system supports the full TPC-H benchmark: the papers state that TQP can run all 22 TPC-H queries (Asada et al., 2022, He et al., 2022).

The principal claim is architectural reuse of the ML ecosystem. By compiling SQL to tensor programs, TQP inherits heterogeneous backend support, profiling, graph visualization, and end-to-end integration with ML operators. The 2022 full paper reports that TQP can improve query execution time by up to 10$$\mathrm{MOS}_T=(\alpha\cdot \mathrm{TC}+\beta)\cdot \mathrm{TQP}+b.$$1 over specialized CPU- and GPU-only systems, and up to 9$$\mathrm{MOS}_T=(\alpha\cdot \mathrm{TC}+\beta)\cdot \mathrm{TQP}+b.$$2 for mixed SQL+ML prediction queries (He et al., 2022). The demonstration paper emphasizes that TQP provides performance comparable to, and often better than, specialized CPU and GPU query processors while also integrating with Pandas for ingestion and TensorBoard for visualization (Asada et al., 2022).

5. Quadratic optimization and beamforming

In multi-user massive-MIMO downlink, TQP denotes Trace Quotient Problem. The outer-beamformer design in two-stage beamforming is derived from a lower bound on the average SLNR under zero-forcing inner beamforming and reduces to

$$\mathrm{MOS}_T=(\alpha\cdot \mathrm{TC}+\beta)\cdot \mathrm{TQP}+b.$$3

where $$\mathrm{MOS}_T=(\alpha\cdot \mathrm{TC}+\beta)\cdot \mathrm{TQP}+b.$$4 encodes effective desired-signal covariance and $$\mathrm{MOS}_T=(\alpha\cdot \mathrm{TC}+\beta)\cdot \mathrm{TQP}+b.$$5 encodes inter-group interference plus noise (Kim et al., 2014). The paper emphasizes that this is the same mathematical form often encountered in machine learning. It then presents an iterative eigenvalue-based algorithm: at iteration $$\mathrm{MOS}_T=(\alpha\cdot \mathrm{TC}+\beta)\cdot \mathrm{TQP}+b.$$6, compute the current trace ratio $$\mathrm{MOS}_T=(\alpha\cdot \mathrm{TC}+\beta)\cdot \mathrm{TQP}+b.$$7, form $$\mathrm{MOS}_T=(\alpha\cdot \mathrm{TC}+\beta)\cdot \mathrm{TQP}+b.$$8, and update $$\mathrm{MOS}_T=(\alpha\cdot \mathrm{TC}+\beta)\cdot \mathrm{TQP}+b.$$9 using the dominant eigenvectors. The objective value is shown to increase monotonically, and the cited TQP properties imply convergence to the global optimum (Kim et al., 2014). In this setting, TQP is not an implementation block but the central variational formulation.

In integer optimization, TQP denotes Ternary Quadratic Problem, namely

$D_G(\mathrm{tNSL})$0

The 2026 paper introduces a dedicated ternary semidefinite programming framework for this problem class, based on the lifted matrix

$D_G(\mathrm{tNSL})$1

with $D_G(\mathrm{tNSL})$2 in the exact model (Meijer et al., 30 Mar 2026). The proposed exact approach strengthens a basic SDP relaxation by triangle, RLT, split, and $D_G(\mathrm{tNSL})$3-gonal inequalities and embeds them in a tailored branch-and-bound algorithm with ternary branching and a feasible-solution heuristic. Computationally, the paper reports that its SDP-B&B solves all tested QUTO and TQP-Linear instances for $D_G(\mathrm{tNSL})$4 within the time limit, whereas GUROBI fails on multiple instances in all three tested instance families; for larger instances, remaining gaps are generally small (Meijer et al., 30 Mar 2026).

These two meanings share ratio-form quadratic structure but occur at different levels. The trace quotient problem is a continuous subspace optimization over orthonormal beamformers, whereas the ternary quadratic problem is a discrete optimization over $D_G(\mathrm{tNSL})$5 with exact semidefinite lifting.

6. Comparative perspective and disambiguation

The cited literature shows that TQP is not a stable cross-disciplinary term. It may refer to a single integer-valued encoder parameter (Long et al., 2024), a structured evaluator prompt (Sato, 16 Apr 2025), an attention-based temporal projection module (Ye et al., 13 Nov 2025), a reduced-reference prediction task (Menon et al., 2023), a query execution system (Asada et al., 2022, He et al., 2022), a trace-ratio optimization problem (Kim et al., 2014), an integer quadratic problem class (Meijer et al., 30 Mar 2026), or a training strategy for sparse MoEs (Shravan, 5 Jun 2026).

The local lexical environment resolves the ambiguity. In compression and PCQA, TQP appears with TBPP, TC, MOS, and tNSL (Long et al., 2024). In LLM prompting, it appears with TIP, Tone Phase, and Tsun-Dere score (Sato, 16 Apr 2025). In action recognition, it appears with Q-Former, skeleton representation, and LLaMA2 (Ye et al., 13 Nov 2025). In streaming, it appears with TQPM, PSNR, SSIM, and VMAF (Menon et al., 2023). In database systems, it appears with tensor computation runtimes, TPC-H, and TorchScript/ONNX/TVM (Asada et al., 2022, He et al., 2022). In optimization and communications, it appears with SLNR, outer beamformer, SDP, and $D_G(\mathrm{tNSL})$6 lifting (Kim et al., 2014, Meijer et al., 30 Mar 2026). In large-scale MoE training, it appears with TurboQuant, rank-16 adapters, and optimizer state (Shravan, 5 Jun 2026).

A plausible implication is that “TQP” should be read as a field-local shorthand rather than a portable concept. In contemporary arXiv usage, its meaning is determined not by the acronym itself but by the surrounding formalism, system model, and notation.