Sherry: Quantization, Caching & Quasar Survey

• Sherry is a hardware-efficient ternary quantization framework that reduces bit-width to 1.25 bits per weight using a 3:4 sparsity scheme, enhancing efficiency in large language models.
• Sherry advances online caching theory by introducing a delayed-hits model and superphase analysis, providing an O(Zk) competitive bound for algorithms like LRU.
• SHERRY is a submillimetre survey of z∼6 quasars using SCUBA2, linking FIR luminosity and star formation rates with weak ultraviolet emission lines in early AGN–galaxy evolution.

Sherry refers to several distinct, influential concepts and research frameworks within modern computational and astronomical literature. It is used as (1) a hardware-efficient ternary quantization framework for neural networks, (2) a key analytical innovator in online caching theory, and (3) the acronym "SHERRY" for a landmark submillimetre survey of high-redshift quasars. Each domain is outlined below with rigor, contextualization, and linkage to research priorities.

1. Sherry: Hardware-Efficient Ternary Quantization Framework

Sherry is a ternary quantization method designed to reduce memory/computation bottlenecks in LLM deployment on resource-constrained hardware. It achieves a hardware-optimal bit-width of 1.25 bits per network weight using a fine-grained 3:4 sparsity scheme and introduces mechanistic solutions for training pathologies unique to structured ternary constraints (Huang et al., 12 Jan 2026).

Formalization: 3:4 Fine-Grained Sparsity

The principal insight is to enforce, for each block of size 4 in the weight matrix $W \in \mathbb{R}^{d_{\rm in} \times d_{\rm out}}$, a ternary approximation $T_{i,j} \in \{-1,0,1\}$ such that exactly 3 of every 4 weights in each block are nonzero. For each column, the quantization objective is:

$$\min_{T_{:,j},\,\alpha_j} \|W_{:,j} - T_{:,j}\alpha_j\|_2^2$$

subject to $T_{i,j} \in \{-1, 0, +1\}$ and $\|T_{b:b+3,j}\|_0 = 3$ for all $b$ (block indices). The optimal block ternarization (Sparse-AbsMean) is constructed by zeroing the weight with the smallest magnitude in each block, setting the remaining three to $\pm1$, and averaging their absolute value for the scaling factor $\alpha_j$.

Each block’s 32 ternary patterns ($\binom{4}{3} \times 2^3 = 32$) fit precisely in 5 bits—thus, when storing 4 weights into a 5-bit word, the effective bit-width is $1.25$ with perfect alignment to 128-bit SIMD instructions.

Weight Trapping Phenomenon

Direct QAT under this ternary sparsification results in "gradient homogenization." The backward-pass gradient through the straight-through estimator (STE) exhibits strongly reduced rank due to the uniform sparsity mask, leading to "representational collapse": model parameters are driven toward binary values, and the ternary codebook flexibility is unexploited, with a measurable loss in expressivity.

Arenas: Annealing Residual Synapse Mechanism

To counteract weight trapping, Sherry augments each quantized linear layer during training with a bypass branch carrying full-precision weights, scaled by an annealing coefficient $T_{i,j} \in \{-1,0,1\}$0 over the training schedule. The forward pass is:

$T_{i,j} \in \{-1,0,1\}$1

The resulting backward gradients propagate both ternary and full-precision signal until late-stage annealing, maintaining high Effective Rank and preventing collapse. At inference, only the ternary branch remains, ensuring zero-cost for the Arenas solution.

Empirical and Hardware Efficiency Results

On LLaMA-3.2 (1B, 3B parameters), Sherry matches or exceeds the leading 1.67-bit schemes (e.g., TL2) in accuracy across ARC-Easy, ARC-Challenge, HellaSwag, PIQA, and WinoGrande, while affording a $T_{i,j} \in \{-1,0,1\}$2–$T_{i,j} \in \{-1,0,1\}$3 bit-width reduction over 2-bit and 1.67-bit baselines. Inference throughput improves by $T_{i,j} \in \{-1,0,1\}$4–$T_{i,j} \in \{-1,0,1\}$5 on consumer Intel CPUs; model memory usage shrinks proportionately. Sherry's packing/unpacking uses routine AVX2 vector instructions and local LUTs, requiring no exotic hardware (Huang et al., 12 Jan 2026).

2. Sherry in Online Caching Theory: Delayed-Hits Model

In the context of online algorithms, "Sherry" denotes a principal contributor to the theoretical analysis of the delayed-hits caching model, a generalization of classical paging where the latency to accommodate a miss is parameterized by a delay $T_{i,j} \in \{-1,0,1\}$6 (Gurushankar et al., 27 Jan 2025).

Delayed-Hits Model Definition

Given universe size $T_{i,j} \in \{-1,0,1\}$7, cache size $T_{i,j} \in \{-1,0,1\}$8, and delay window $T_{i,j} \in \{-1,0,1\}$9, requests $$\min_{T_{:,j},\,\alpha_j} \|W_{:,j} - T_{:,j}\alpha_j\|_2^2$$0 are served over $$\min_{T_{:,j},\,\alpha_j} \|W_{:,j} - T_{:,j}\alpha_j\|_2^2$$1 timesteps. Fetches take $$\min_{T_{:,j},\,\alpha_j} \|W_{:,j} - T_{:,j}\alpha_j\|_2^2$$2 steps to complete; a request during an in-flight fetch may incur reduced penalty ("delayed-hit"). The instantaneous cost $$\min_{T_{:,j},\,\alpha_j} \|W_{:,j} - T_{:,j}\alpha_j\|_2^2$$3 is:

• $$\min_{T_{:,j},\,\alpha_j} \|W_{:,j} - T_{:,j}\alpha_j\|_2^2$$4 (cache hit)
• $$\min_{T_{:,j},\,\alpha_j} \|W_{:,j} - T_{:,j}\alpha_j\|_2^2$$5 (hit during delay window)
• $$\min_{T_{:,j},\,\alpha_j} \|W_{:,j} - T_{:,j}\alpha_j\|_2^2$$6 (miss, no in-flight fetch)

The goal is to minimize total latency $$\min_{T_{:,j},\,\alpha_j} \|W_{:,j} - T_{:,j}\alpha_j\|_2^2$$7.

$$\min_{T_{:,j},\,\alpha_j} \|W_{:,j} - T_{:,j}\alpha_j\|_2^2$$8 Competitiveness of LRU

Sherry, with collaborators, developed the "superphase" analysis, refining classical phase partitioning to group phases into superphases of length at least $$\min_{T_{:,j},\,\alpha_j} \|W_{:,j} - T_{:,j}\alpha_j\|_2^2$$9. Their key result is that Least Recently Used (LRU) and more generally any "marking algorithm" incurs at most $T_{i,j} \in \{-1, 0, +1\}$0 times the optimal offline cost:

$T_{i,j} \in \{-1, 0, +1\}$1

The proof leverages:

• Phase and superphase decomposition
• Marking invariants (a requested page remains persistent in cache throughout the phase)
• Tight bounding of LRU cost within superphases
• Matching lower-bound constructions

For $T_{i,j} \in \{-1, 0, +1\}$2, the result recovers Sleator–Tarjan's $T_{i,j} \in \{-1, 0, +1\}$3 bound for classical paging.

Novel Techniques

Sherry's introduction of superphase decomposition generalizes competitive analysis to a pipelined delay model, enabling tight asymptotics in $T_{i,j} \in \{-1, 0, +1\}$4 and $T_{i,j} \in \{-1, 0, +1\}$5. The analysis applies directly to marking-style algorithms and paves the way for further exploration of delayed-service models.

Prospective Directions

Subsequent research directions include incorporating prediction, handling weighted or variable-sized pages, exploiting parallel fetches ($T_{i,j} \in \{-1, 0, +1\}$6), and robustness to fluctuating or uncertain $T_{i,j} \in \{-1, 0, +1\}$7. The framework is positioned as broadly applicable to resource-lockin phenomena in online scheduling (Gurushankar et al., 27 Jan 2025).

3. SHERRY: SCUBA2 High Redshift Bright Quasar Survey

SHERRY (SCUBA2 High rEdshift bRight quasaR surveY) is a submillimeter continuum survey targeting the far-infrared properties and spectral signatures of $T_{i,j} \in \{-1, 0, +1\}$8 quasars using the SCUBA2 instrument on JCMT (Li et al., 2020).

Survey Design and Objectives

• Sample: 54 optically/NIR-detected quasars with $T_{i,j} \in \{-1, 0, +1\}$9, $\|T_{b:b+3,j}\|_0 = 3$0, avoiding duplication with previous mm/submm surveys.
• Instrument: SCUBA2 at 450 $\|T_{b:b+3,j}\|_0 = 3$1m and 850 $\|T_{b:b+3,j}\|_0 = 3$2m, with $\|T_{b:b+3,j}\|_0 = 3$31.2 mJy beam$\|T_{b:b+3,j}\|_0 = 3$4 rms at 850 $\|T_{b:b+3,j}\|_0 = 3$5m.
• Detection Criterion: $\|T_{b:b+3,j}\|_0 = 3$6 at source position.
• Goals:
  • Quantify FIR luminosity ($\|T_{b:b+3,j}\|_0 = 3$7) and dust-continuum emission.
  • Derive star formation rates (SFRs) and dust masses in host galaxies.
  • Systematically analyze weak-line quasar (WLQ) incidence ($\|T_{b:b+3,j}\|_0 = 3$8(Ly$\|T_{b:b+3,j}\|_0 = 3$9+Nv)$b$0Å).

Key Observational Results

• Detection Rate: 16/54 (30%) have secure $b$1m detections; median flux $b$2 mJy for detections.
• FIR/SFR: $b$3–$b$4; SFR$b$5–$b$6 yr$b$7 inferred from greybody dust models ($b$8 K, $b$9).
• Comparison: At $\pm1$0, ultra-luminous FIR hosts ($\pm1$1) are rarer than in $\pm1$2–$\pm1$3 samples.

Ultraviolet Spectral Diagnostics

• Weak-Line Incidence: 11% (6/54) classified as WLQs ($\pm1$415.4 Å).
• Detections have systematically lower $\pm1$5: mean $\pm1$6 ($\pm1$722 Å) versus $\pm1$8 ($\pm1$960 Å) for non-detections; K–S test $\alpha_j$0.
• This suggests a statistically significant link between strong dust emission and weak UV emission lines.

Interpretive Framework

The SHERRY results are consistent with two scenarios for coevolving AGN and host: (a) extremely high Eddington ratio accretion creating a thick "shielding gas" structure, with UV lines suppressed and ISM fueling intense starbursts; or (b) evolutionary phases with underdeveloped broad line regions and high SFRs during rapid black hole and stellar mass assembly.

A plausible implication is that SHERRY is probing an early, formative AGN–galaxy phase at cosmic dawn—high SFRs, strong dust continuum, and weak-line regions that reflect the interplay between accretion geometry and ISM conditions (Li et al., 2020).

4. Summary Table: Key Dimensions of "Sherry"

Context	Domain & Purpose	Foundational Reference
Quantization Framework	Hardware-efficient, 1.25-bit ternary LLM quantization	(Huang et al., 12 Jan 2026)
Caching Theory	Analysis of delayed-hits model, $\alpha_j$1 competitive LRU bounds	(Gurushankar et al., 27 Jan 2025)
Astronomy Survey	SHERRY: Submm continuum/properties of $\alpha_j$2 quasars	(Li et al., 2020)

5. Research Impact and Future Directions

In quantized neural inference, Sherry establishes the practical bit-width lower bound for ternary models compatible with modern SIMD hardware, addresses training instabilities, and demonstrates scalable performance on LLMs up to 3B parameters. Extending to even larger models, integrating activation quantization, and leveraging sparse tensor cores remain open challenges (Huang et al., 12 Jan 2026).

In online caching, Sherry's analytical techniques for the delayed-hits setting provide the field's first matching $\alpha_j$3 guarantees. Future work will likely explore augmentations with predictions, cost-scaling in more complex resource landscapes, and generalization to multi-channel or dynamic-delay architectures (Gurushankar et al., 27 Jan 2025).

As a survey, SHERRY robustly characterizes both the incidence of starbursting hosts and their correspondence to weak emission lines, informing AGN–galaxy coevolutionary models in the early universe. Additional spectral follow-up and extension to fainter luminosities or higher redshifts are prospective avenues (Li et al., 2020).